Analytics In A Big Data World Bart Baesens The Essential Guide To Science And Its Applicati

Bart%20Baesens%20Analytics%20in%20a%20Big%20Data%20World%20The%20Essential%20Guide%20to%20Data%20Science%20and%20its%20Applicati

User Manual:

Open the PDF directly: View PDF .
Page Count: 252

Download
Open PDF In Browser	View PDF

Analytics in a Big
Data World

Wiley & SAS
Business Series

The Wiley & SAS Business Series presents books that help senior‐level
managers with their critical management decisions.
Titles in the Wiley & SAS Business Series include:
Activity‐Based Management for Financial Institutions: Driving Bottom‐
Line Results by Brent Bahnub
Bank Fraud: Using Technology to Combat Losses by Revathi Subramanian
Big Data Analytics: Turning Big Data into Big Money by Frank Ohlhorst
Branded! How Retailers Engage Consumers with Social Media and Mobility by Bernie Brennan and Lori Schafer
Business Analytics for Customer Intelligence by Gert Laursen
Business Analytics for Managers: Taking Business Intelligence beyond
Reporting by Gert Laursen and Jesper Thorlund
The Business Forecasting Deal: Exposing Bad Practices and Providing
Practical Solutions by Michael Gilliland
Business Intelligence Applied: Implementing an Effective Information and
Communications Technology Infrastructure by Michael Gendron
Business Intelligence in the Cloud: Strategic Implementation Guide by
Michael S. Gendron
Business Intelligence Success Factors: Tools for Aligning Your Business in
the Global Economy by Olivia Parr Rud
CIO Best Practices: Enabling Strategic Value with Information Technology,
second edition by Joe Stenzel
Connecting Organizational Silos: Taking Knowledge Flow Management to
the Next Level with Social Media by Frank Leistner
Credit Risk Assessment: The New Lending System for Borrowers, Lenders,
and Investors by Clark Abrahams and Mingyuan Zhang

Credit Risk Scorecards: Developing and Implementing Intelligent Credit
Scoring by Naeem Siddiqi
The Data Asset: How Smart Companies Govern Their Data for Business
Success by Tony Fisher
Delivering Business Analytics: Practical Guidelines for Best Practice by
Evan Stubbs
Demand‐Driven Forecasting: A Structured Approach to Forecasting, Second Edition by Charles Chase
Demand‐Driven Inventory Optimization and Replenishment: Creating a
More Efficient Supply Chain by Robert A. Davis
The Executive’s Guide to Enterprise Social Media Strategy: How Social Networks Are Radically Transforming Your Business by David Thomas and
Mike Barlow
Economic and Business Forecasting: Analyzing and Interpreting Econometric Results by John Silvia, Azhar Iqbal, Kaylyn Swankoski, Sarah
Watt, and Sam Bullard
Executive’s Guide to Solvency III by David Buckham, Jason Wahl, and
Stuart Rose
Fair Lending Compliance: Intelligence and Implications for Credit Risk
Managementt by Clark R. Abrahams and Mingyuan Zhang
Foreign Currency Financial Reporting from Euros to Yen to Yuan: A Guide
to Fundamental Concepts and Practical Applications by Robert Rowan
Health Analytics: Gaining the Insights to Transform Health Care by Jason
Burke
Heuristics in Analytics: A Practical Perspective of What Influences Our
Analytical World
d by Carlos Andre Reis Pinheiro and Fiona McNeill
Human Capital Analytics: How to Harness the Potential of Your Organization’s Greatest Assett by Gene Pease, Boyce Byerly, and Jac Fitz‐enz
Implement, Improve and Expand Your Statewide Longitudinal Data System: Creating a Culture of Data in Education by Jamie McQuiggan and
Armistead Sapp
Information Revolution: Using the Information Evolution Model to Grow
Your Business by Jim Davis, Gloria J. Miller, and Allan Russell

Killer Analytics: Top 20 Metrics Missing from Your Balance Sheett by Mark
Brown
Manufacturing Best Practices: Optimizing Productivity and Product Quality by Bobby Hull
Marketing Automation: Practical Steps to More Effective Direct Marketing
by Jeff LeSueur
Mastering Organizational Knowledge Flow: How to Make Knowledge
Sharing Workk by Frank Leistner
The New Know: Innovation Powered by Analytics by Thornton May
Performance Management: Integrating Strategy Execution, Methodologies,
Risk, and Analytics by Gary Cokins
Predictive Business Analytics: Forward‐Looking Capabilities to Improve
Business Performance by Lawrence Maisel and Gary Cokins
Retail Analytics: The Secret Weapon by Emmett Cox
Social Network Analysis in Telecommunications by Carlos Andre Reis
Pinheiro
Statistical Thinking: Improving Business Performance, second edition by
Roger W. Hoerl and Ronald D. Snee
Taming the Big Data Tidal Wave: Finding Opportunities in Huge Data
Streams with Advanced Analytics by Bill Franks
Too Big to Ignore: The Business Case for Big Data by Phil Simon
The Value of Business Analytics: Identifying the Path to Profitability by
Evan Stubbs
Visual Six Sigma: Making Data Analysis Lean by Ian Cox, Marie A.
Gaudard, Philip J. Ramsey, Mia L. Stephens, and Leo Wright
Win with Advanced Business Analytics: Creating Business Value from
Your Data by Jean Paul Isson and Jesse Harriott
For more information on any of the above titles, please visit www
.wiley.com.

Analytics in a Big
Data World
The Essential Guide to Data Science
and Its Applications

Bart Baesens

Cover image: ©iStockphoto/vlastos
Cover design: Wiley
Copyright © 2014 by Bart Baesens. All rights reserved.
Published by John Wiley & Sons, Inc., Hoboken, New Jersey.
Published simultaneously in Canada.
No part of this publication may be reproduced, stored in a retrieval system,
or transmitted in any form or by any means, electronic, mechanical,
photocopying, recording, scanning, or otherwise, except as permitted under
Section 107 or 108 of the 1976 United States Copyright Act, without either the
prior written permission of the Publisher, or authorization through payment
of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222
Rosewood Drive, Danvers, MA 01923, (978) 750-8400, fax (978) 646-8600, or
on the Web at www.copyright.com. Requests to the Publisher for permission
should be addressed to the Permissions Department, John Wiley & Sons, Inc.,
111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or
online at http://www.wiley.com/go/permissions.
Limit of Liability/Disclaimer of Warranty: While the publisher and author have
used their best efforts in preparing this book, they make no representations
or warranties with respect to the accuracy or completeness of the contents of
this book and specifically disclaim any implied warranties of merchantability
or fitness for a particular purpose. No warranty may be created or extended
by sales representatives or written sales materials. The advice and strategies
contained herein may not be suitable for your situation. You should consult
with a professional where appropriate. Neither the publisher nor author shall
be liable for any loss of profit or any other commercial damages, including but
not limited to special, incidental, consequential, or other damages.
For general information on our other products and services or for technical
support, please contact our Customer Care Department within the United
States at (800) 762-2974, outside the United States at (317) 572-3993 or
fax (317) 572-4002.
Wiley publishes in a variety of print and electronic formats and by print-ondemand. Some material included with standard print versions of this book
may not be included in e-books or in print-on-demand. If this book refers to
media such as a CD or DVD that is not included in the version you purchased,
you may download this material at http://booksupport.wiley.com. For more
information about Wiley products, visit www.wiley.com.
Library of Congress Cataloging-in-Publication Data:
Baesens, Bart.
Analytics in a big data world : the essential guide to data science and its
applications / Bart Baesens.
1 online resource. — (Wiley & SAS business series)
Description based on print version record and CIP data provided by publisher;
resource not viewed.
ISBN 978-1-118-89271-8 (ebk); ISBN 978-1-118-89274-9 (ebk);
ISBN 978-1-118-89270-1 (cloth) 1. Big data. 2. Management—Statistical
methods. 3. Management—Data processing. 4. Decision making—Data
processing. I. Title.
HD30.215
658.4’038 dc23
2014004728
Printed in the United States of America
10 9 8 7 6 5 4 3 2 1

To my wonderful wife, Katrien, and my kids,
Ann-Sophie, Victor, and Hannelore.
To my parents and parents-in-law.

Contents

Preface

xiii

Acknowledgments xv
Chapter 1

Big Data and Analytics 1

Example Applications 2
Basic Nomenclature 4
Analytics Process Model
Job Profiles Involved
Analytics

Analytical Model Requirements
Notes

Chapter 2 Data Collection, Sampling,
and Preprocessing 13
Types of Data Sources
Sampling

Types of Data Elements

Visual Data Exploration and Exploratory
Statistical Analysis 17
Missing Values

Outlier Detection and Treatment
Standardizing Data

Categorization 24
Weights of Evidence Coding
Variable Selection

▸ C ON T E N T S

Segmentation
Notes
Chapter 3

33
Predictive Analytics

Target Definition

Linear Regression

Logistic Regression

Decision Trees 42
Neural Networks

Support Vector Machines

Ensemble Methods 64
Multiclass Classification Techniques

Evaluating Predictive Models 71
Notes

Chapter 4 Descriptive Analytics
Association Rules
Sequence Rules
Segmentation
Notes
Chapter 5

87
94

104
Survival Analysis

105

Survival Analysis Measurements

106

Kaplan Meier Analysis 109
Parametric Survival Analysis 111
Proportional Hazards Regression

114

Extensions of Survival Analysis Models 116
Evaluating Survival Analysis Models 117
Notes
Chapter 6

117
Social Network Analytics

Social Network Definitions

119

Social Network Metrics 121
Social Network Learning

123

Relational Neighbor Classifier

124

119

CONTENTS ◂

Probabilistic Relational Neighbor Classifier
Relational Logistic Regression
Collective Inferencing
Egonets

125

126

128

129

Bigraphs 130
Notes

132

Chapter 7 Analytics: Putting It All to Work
Backtesting Analytical Models 134
Benchmarking 146
Data Quality

149

Software 153
Privacy

155

Model Design and Documentation 158
Corporate Governance
Notes

159

Chapter 8

Example Applications

Credit Risk Modeling

161

Fraud Detection 165
Net Lift Response Modeling 168
Churn Prediction

172

Recommender Systems 176
Web Analytics

185

Social Media Analytics 195
Business Process Analytics 204
Notes

220

About the Author
Index

225

223

161

133

Preface

ompanies are being flooded with tsunamis of data collected in a
multichannel business environment, leaving an untapped potential for analytics to better understand, manage, and strategically
exploit the complex dynamics of customer behavior. In this book, we
will discuss how analytics can be used to create strategic leverage and
identify new business opportunities.
The focus of this book is not on the mathematics or theory, but on
the practical application. Formulas and equations will only be included
when absolutely needed from a practitioner’s perspective. It is also not
our aim to provide exhaustive coverage of all analytical techniques
previously developed, but rather to cover the ones that really provide
added value in a business setting.
The book is written in a condensed, focused way because it is targeted at the business professional. A reader’s prerequisite knowledge
should consist of some basic exposure to descriptive statistics (e.g.,
mean, standard deviation, correlation, confidence intervals, hypothesis
testing), data handling (using, for example, Microsoft Excel, SQL, etc.),
and data visualization (e.g., bar plots, pie charts, histograms, scatter
plots). Throughout the book, many examples of real‐life case studies
will be included in areas such as risk management, fraud detection,
customer relationship management, web analytics, and so forth. The
author will also integrate both his research and consulting experience
throughout the various chapters. The book is aimed at senior data analysts, consultants, analytics practitioners, and PhD researchers starting
to explore the field.
Chapter 1 discusses big data and analytics. It starts with some
example application areas, followed by an overview of the analytics
process model and job profiles involved, and concludes by discussing
key analytic model requirements. Chapter 2 provides an overview of
xiii

xiv

▸ P R E FAC E

data collection, sampling, and preprocessing. Data is the key ingredient to any analytical exercise, hence the importance of this chapter.
It discusses sampling, types of data elements, visual data exploration
and exploratory statistical analysis, missing values, outlier detection
and treatment, standardizing data, categorization, weights of evidence
coding, variable selection, and segmentation. Chapter 3 discusses predictive analytics. It starts with an overview of the target definition
and then continues to discuss various analytics techniques such as
linear regression, logistic regression, decision trees, neural networks,
support vector machines, and ensemble methods (bagging, boosting, random forests). In addition, multiclass classification techniques
are covered, such as multiclass logistic regression, multiclass decision trees, multiclass neural networks, and multiclass support vector
machines. The chapter concludes by discussing the evaluation of predictive models. Chapter 4 covers descriptive analytics. First, association
rules are discussed that aim at discovering intratransaction patterns.
This is followed by a section on sequence rules that aim at discovering
intertransaction patterns. Segmentation techniques are also covered.
Chapter 5 introduces survival analysis. The chapter starts by introducing some key survival analysis measurements. This is followed by a
discussion of Kaplan Meier analysis, parametric survival analysis, and
proportional hazards regression. The chapter concludes by discussing
various extensions and evaluation of survival analysis models. Chapter 6 covers social network analytics. The chapter starts by discussing
example social network applications. Next, social network definitions
and metrics are given. This is followed by a discussion on social network
learning. The relational neighbor classifier and its probabilistic variant
together with relational logistic regression are covered next. The chapter ends by discussing egonets and bigraphs. Chapter 7 provides an
overview of key activities to be considered when putting analytics to
work. It starts with a recapitulation of the analytic model requirements
and then continues with a discussion of backtesting, benchmarking,
data quality, software, privacy, model design and documentation, and
corporate governance. Chapter 8 concludes the book by discussing various example applications such as credit risk modeling, fraud detection,
net lift response modeling, churn prediction, recommender systems,
web analytics, social media analytics, and business process analytics.

Acknowledgments

would like to acknowledge all my colleagues who contributed to
this text: Seppe vanden Broucke, Alex Seret, Thomas Verbraken,
Aimée Backiel, Véronique Van Vlasselaer, Helen Moges, and Barbara
Dergent.

Analytics in a Big
Data World

C H A P T E R

Big Data and
Analytics

ata are everywhere. IBM projects that every day we generate 2.5
quintillion bytes of data.1 In relative terms, this means 90 percent
of the data in the world has been created in the last two years.
Gartner projects that by 2015, 85 percent of Fortune 500 organizations
will be unable to exploit big data for competitive advantage and about
4.4 million jobs will be created around big data.2 Although these estimates should not be interpreted in an absolute sense, they are a strong
indication of the ubiquity of big data and the strong need for analytical
skills and resources because, as the data piles up, managing and analyzing these data resources in the most optimal way become critical success factors in creating competitive advantage and strategic leverage.
Figure 1.1 shows the results of a KDnuggets3 poll conducted during April 2013 about the largest data sets analyzed. The total number
of respondents was 322 and the numbers per category are indicated
between brackets. The median was estimated to be in the 40 to 50 gigabyte (GB) range, which was about double the median answer for a similar poll run in 2012 (20 to 40 GB). This clearly shows the quick increase
in size of data that analysts are working on. A further regional breakdown of the poll showed that U.S. data miners lead other regions in big
data, with about 28% of them working with terabyte (TB) size databases.
A main obstacle to fully harnessing the power of big data using analytics is the lack of skilled resources and “data scientist” talent required to
1

▸ A N A LY T I C S I N A BI G DATA WO RL D

3.7%

Less than 1 MB (12)
1.1 to 10 MB (8)

2.5%
4.3%

11 to 100 MB (14)

15.5%

101 MB to 1 GB (50)
1.1 to 10 GB (59)

18%
16%

11 to 100 GB (52)
101 GB to 1 TB
(59)

18%
12%

1.1 to 10 TB (39)
4.7%

11 to 100 TB (15)
101 TB to 1 PB (6)

1.9%

1.1 to 10 PB (2)

0.6%

11 to 100 PB (0)

Over 100 PB (6)

1.9%

Figure 1.1 Results from a KDnuggets Poll about Largest Data Sets Analyzed
Source: www.kdnuggets.com/polls/2013/largest‐dataset‐analyzed‐data‐mined‐2013.html.

exploit big data. In another poll ran by KDnuggets in July 2013, a strong
need emerged for analytics/big data/data mining/data science education.4 It is the purpose of this book to try and fill this gap by providing a
concise and focused overview of analytics for the business practitioner.

EXAMPLE APPLICATIONS
Analytics is everywhere and strongly embedded into our daily lives. As I
am writing this part, I was the subject of various analytical models today.
When I checked my physical mailbox this morning, I found a catalogue
sent to me most probably as a result of a response modeling analytical
exercise that indicated that, given my characteristics and previous purchase behavior, I am likely to buy one or more products from it. Today,
I was the subject of a behavioral scoring model of my financial institution. This is a model that will look at, among other things, my checking account balance from the past 12 months and my credit payments
during that period, together with other kinds of information available
to my bank, to predict whether I will default on my loan during the
next year. My bank needs to know this for provisioning purposes. Also
today, my telephone services provider analyzed my calling behavior

B I G D ATA A N D A N A LY T I C S ◂

and my account information to predict whether I will churn during the
next three months. As I logged on to my Facebook page, the social ads
appearing there were based on analyzing all information (posts, pictures,
my friends and their behavior, etc.) available to Facebook. My Twitter
posts will be analyzed (possibly in real time) by social media analytics to
understand both the subject of my tweets and the sentiment of them.
As I checked out in the supermarket, my loyalty card was scanned first,
followed by all my purchases. This will be used by my supermarket to
analyze my market basket, which will help it decide on product bundling, next best offer, improving shelf organization, and so forth. As I
made the payment with my credit card, my credit card provider used
a fraud detection model to see whether it was a legitimate transaction.
When I receive my credit card statement later, it will be accompanied by
various vouchers that are the result of an analytical customer segmentation exercise to better understand my expense behavior.
To summarize, the relevance, importance, and impact of analytics
are now bigger than ever before and, given that more and more data
are being collected and that there is strategic value in knowing what
is hidden in data, analytics will continue to grow. Without claiming to
be exhaustive, Table 1.1 presents some examples of how analytics is
applied in various settings.
Table 1.1

Example Analytics Applications

Marketing

Risk
Management Government Web

Response
modeling

Credit risk
modeling

Tax avoidance Web analytics Demand
forecasting

Net lift
modeling

Market risk
modeling

Social
security fraud

Social media
analytics

Retention
modeling

Operational
risk modeling

Money
laundering

Multivariate
testing

Market basket
analysis

Fraud
detection

Terrorism
detection

Recommender
systems
Customer
segmentation

Logistics

Other
Text
analytics

Supply chain Business
analytics
process
analytics

▸ A N A LY T I C S I N A BI G DATA WO RL D

It is the purpose of this book to discuss the underlying techniques
and key challenges to work out the applications shown in Table 1.1
using analytics. Some of these applications will be discussed in further
detail in Chapter 8.

BASIC NOMENCLATURE
In order to start doing analytics, some basic vocabulary needs to be
defined. A first important concept here concerns the basic unit of analysis. Customers can be considered from various perspectives. Customer
lifetime value (CLV) can be measured for either individual customers
or at the household level. Another alternative is to look at account
behavior. For example, consider a credit scoring exercise for which
the aim is to predict whether the applicant will default on a particular
mortgage loan account. The analysis can also be done at the transaction level. For example, in insurance fraud detection, one usually performs the analysis at insurance claim level. Also, in web analytics, the
basic unit of analysis is usually a web visit or session.
It is also important to note that customers can play different roles.
For example, parents can buy goods for their kids, such that there is
a clear distinction between the payer and the end user. In a banking
setting, a customer can be primary account owner, secondary account
owner, main debtor of the credit, codebtor, guarantor, and so on. It
is very important to clearly distinguish between those different roles
when defining and/or aggregating data for the analytics exercise.
Finally, in case of predictive analytics, the target variable needs to
be appropriately defined. For example, when is a customer considered
to be a churner or not, a fraudster or not, a responder or not, or how
should the CLV be appropriately defined?

ANALYTICS PROCESS MODEL
Figure 1.2 gives a high‐level overview of the analytics process model.5
As a first step, a thorough definition of the business problem to be
solved with analytics is needed. Next, all source data need to be identified that could be of potential interest. This is a very important step, as
data is the key ingredient to any analytical exercise and the selection of

B I G D ATA A N D A N A LY T I C S ◂

data will have a deterministic impact on the analytical models that will
be built in a subsequent step. All data will then be gathered in a staging area, which could be, for example, a data mart or data warehouse.
Some basic exploratory analysis can be considered here using, for
example, online analytical processing (OLAP) facilities for multidimensional data analysis (e.g., roll‐up, drill down, slicing and dicing). This
will be followed by a data cleaning step to get rid of all inconsistencies,
such as missing values, outliers, and duplicate data. Additional transformations may also be considered, such as binning, alphanumeric to
numeric coding, geographical aggregation, and so forth. In the analytics step, an analytical model will be estimated on the preprocessed and
transformed data. Different types of analytics can be considered here
(e.g., to do churn prediction, fraud detection, customer segmentation,
market basket analysis). Finally, once the model has been built, it will
be interpreted and evaluated by the business experts. Usually, many
trivial patterns will be detected by the model. For example, in a market
basket analysis setting, one may find that spaghetti and spaghetti sauce
are often purchased together. These patterns are interesting because
they provide some validation of the model. But of course, the key issue
here is to find the unexpected yet interesting and actionable patterns
(sometimes also referred to as knowledge diamonds) that can provide
added value in the business setting. Once the analytical model has
been appropriately validated and approved, it can be put into production as an analytics application (e.g., decision support system, scoring
engine). It is important to consider here how to represent the model
output in a user‐friendly way, how to integrate it with other applications (e.g., campaign management tools, risk engines), and how to
make sure the analytical model can be appropriately monitored and
backtested on an ongoing basis.
It is important to note that the process model outlined in Figure 1.2 is iterative in nature, in the sense that one may have to go back
to previous steps during the exercise. For example, during the analytics step, the need for additional data may be identified, which may
necessitate additional cleaning, transformation, and so forth. Also, the
most time consuming step is the data selection and preprocessing step;
this usually takes around 80% of the total efforts needed to build an
analytical model.

▸ A N A LY T I C S I N A BI G DATA WO RL D

Interpretation
and
Evaluation

Dumps of Operational Data
Understanding
what data is
needed for the
application

Data
Transformation
(binning, alpha to
numeric, etc.)

Data Cleaning
Data
Selection

Patterns
Transformed
Data

Source
Data

Analytics

Data Mining
Mart

Analytics
Application

Preprocessed
Data

Figure 1.2 The Analytics Process Model

JOB PROFILES INVOLVED
Analytics is essentially a multidisciplinary exercise in which many
different job profiles need to collaborate together. In what follows, we
will discuss the most important job profiles.
The database or data warehouse administrator (DBA) is aware of
all the data available within the firm, the storage details, and the data
definitions. Hence, the DBA plays a crucial role in feeding the analytical modeling exercise with its key ingredient, which is data. Because
analytics is an iterative exercise, the DBA may continue to play an
important role as the modeling exercise proceeds.
Another very important profile is the business expert. This could,
for example, be a credit portfolio manager, fraud detection expert,
brand manager, or e‐commerce manager. This person has extensive
business experience and business common sense, which is very valuable. It is precisely this knowledge that will help to steer the analytical
modeling exercise and interpret its key findings. A key challenge here
is that much of the expert knowledge is tacit and may be hard to elicit
at the start of the modeling exercise.
Legal experts are becoming more and more important given that
not all data can be used in an analytical model because of privacy,

B I G D ATA A N D A N A LY T I C S ◂

discrimination, and so forth. For example, in credit risk modeling, one
can typically not discriminate good and bad customers based upon
gender, national origin, or religion. In web analytics, information is
typically gathered by means of cookies, which are files that are stored
on the user’s browsing computer. However, when gathering information using cookies, users should be appropriately informed. This is subject to regulation at various levels (both national and, for example,
European). A key challenge here is that privacy and other regulation
highly vary depending on the geographical region. Hence, the legal
expert should have good knowledge about what data can be used
when, and what regulation applies in what location.
The data scientist, data miner, or data analyst is the person responsible for doing the actual analytics. This person should possess a thorough understanding of all techniques involved and know how to
implement them using the appropriate software. A good data scientist
should also have good communication and presentation skills to report
the analytical findings back to the other parties involved.
The software tool vendors should also be mentioned as an
important part of the analytics team. Different types of tool vendors can
be distinguished here. Some vendors only provide tools to automate
specific steps of the analytical modeling process (e.g., data preprocessing). Others sell software that covers the entire analytical modeling
process. Some vendors also provide analytics‐based solutions for specific application areas, such as risk management, marketing analytics
and campaign management, and so on.

ANALYTICS
Analytics is a term that is often used interchangeably with data science,
data mining, knowledge discovery, and others. The distinction between
all those is not clear cut. All of these terms essentially refer to extracting useful business patterns or mathematical decision models from a
preprocessed data set. Different underlying techniques can be used for
this purpose, stemming from a variety of different disciplines, such as:
■

Statistics (e.g., linear and logistic regression)

■

Machine learning (e.g., decision trees)

▸ A N A LY T I C S I N A BI G DATA WO RL D

■

Biology (e.g., neural networks, genetic algorithms, swarm intelligence)

■

Kernel methods (e.g., support vector machines)

Basically, a distinction can be made between predictive and descriptive analytics. In predictive analytics, a target variable is typically available, which can either be categorical (e.g., churn or not, fraud or not)
or continuous (e.g., customer lifetime value, loss given default). In
descriptive analytics, no such target variable is available. Common
examples here are association rules, sequence rules, and clustering.
Figure 1.3 provides an example of a decision tree in a classification
predictive analytics setting for predicting churn.
More than ever before, analytical models steer the strategic risk
decisions of companies. For example, in a bank setting, the minimum equity and provisions a financial institution holds are directly
determined by, among other things, credit risk analytics, market risk
analytics, operational risk analytics, fraud analytics, and insurance
risk analytics. In this setting, analytical model errors directly affect
profitability, solvency, shareholder value, the macroeconomy, and
society as a whole. Hence, it is of the utmost importance that analytical

Customer Age Recency Frequency Monetary

Churn

John

100

Yes

Sophie

150

Victor

Laura

280

Yes

Age < 40
No

Analytics
Software

Frequency < 5

Recency < 10
Yes

Churn

Figure 1.3

Yes

No Churn

Example of Classification Predictive Analytics

Yes

Churn

No Churn

B I G D ATA A N D A N A LY T I C S ◂

models are developed in the most optimal way, taking into account
various requirements that will be discussed in what follows.

ANALYTICAL MODEL REQUIREMENTS
A good analytical model should satisfy several requirements, depending on the application area. A first critical success factor is business
relevance. The analytical model should actually solve the business
problem for which it was developed. It makes no sense to have a working analytical model that got sidetracked from the original problem
statement. In order to achieve business relevance, it is of key importance that the business problem to be solved is appropriately defined,
qualified, and agreed upon by all parties involved at the outset of the
analysis.
A second criterion is statistical performance. The model should
have statistical significance and predictive power. How this can be measured will depend upon the type of analytics considered. For example,
in a classification setting (churn, fraud), the model should have good
discrimination power. In a clustering setting, the clusters should be as
homogenous as possible. In later chapters, we will extensively discuss
various measures to quantify this.
Depending on the application, analytical models should also be
interpretable and justifiable. Interpretability refers to understanding
the patterns that the analytical model captures. This aspect has a
certain degree of subjectivism, since interpretability may depend on
the business user’s knowledge. In many settings, however, it is considered to be a key requirement. For example, in credit risk modeling
or medical diagnosis, interpretable models are absolutely needed to
get good insight into the underlying data patterns. In other settings,
such as response modeling and fraud detection, having interpretable
models may be less of an issue. Justifiability refers to the degree to
which a model corresponds to prior business knowledge and intuition.6 For example, a model stating that a higher debt ratio results
in more creditworthy clients may be interpretable, but is not justifiable because it contradicts basic financial intuition. Note that both
interpretability and justifiability often need to be balanced against
statistical performance. Often one will observe that high performing

▸ A N A LY T I C S IN A BI G DATA WO RL D

analytical models are incomprehensible and black box in nature.
A popular example of this is neural networks, which are universal
approximators and are high performing, but offer no insight into the
underlying patterns in the data. On the contrary, linear regression
models are very transparent and comprehensible, but offer only
limited modeling power.
Analytical models should also be operationally efficient.
t This refers to
the efforts needed to collect the data, preprocess it, evaluate the model,
and feed its outputs to the business application (e.g., campaign management, capital calculation). Especially in a real‐time online scoring
environment (e.g., fraud detection) this may be a crucial characteristic.
Operational efficiency also entails the efforts needed to monitor and
backtest the model, and reestimate it when necessary.
Another key attention point is the economic costt needed to set up
the analytical model. This includes the costs to gather and preprocess
the data, the costs to analyze the data, and the costs to put the resulting analytical models into production. In addition, the software costs
and human and computing resources should be taken into account
here. It is important to do a thorough cost–benefit analysis at the start
of the project.
Finally, analytical models should also comply with both local and
international regulation and legislation. For example, in a credit risk setting, the Basel II and Basel III Capital Accords have been introduced
to appropriately identify the types of data that can or cannot be used
to build credit risk models. In an insurance setting, the Solvency II
Accord plays a similar role. Given the importance of analytics nowadays, more and more regulation is being introduced relating to the
development and use of the analytical models. In addition, in the context of privacy, many new regulatory developments are taking place at
various levels. A popular example here concerns the use of cookies in
a web analytics context.

NOTES
1.
2.
3.
4.

IBM, www.ibm.com/big‐data/us/en, 2013.
www.gartner.com/technology/topics/big‐data.jsp.
www.kdnuggets.com/polls/2013/largest‐dataset‐analyzed‐data‐mined‐2013.html.
www.kdnuggets.com/polls/2013/analytics‐data‐science‐education.html.

B I G D ATA A N D A N A LY T I C S ◂

5. J. Han and M. Kamber, Data Mining: Concepts and Techniques, 2nd ed. (Morgan
Kaufmann, Waltham, MA, US, 2006); D. J. Hand, H. Mannila, and P. Smyth, Principles of Data Mining (MIT Press, Cambridge, Massachusetts, London, England, 2001);
P. N. Tan, M. Steinbach, and V. Kumar, Introduction to Data Mining (Pearson, Upper
Saddle River, New Jersey, US, 2006).
6. D. Martens, J. Vanthienen, W. Verbeke, and B. Baesens, “Performance of Classification Models from a User Perspective.” Special issue, Decision Support Systems 51, no. 4
(2011): 782–793.

C H A P T E R

Data Collection,
Sampling, and
Preprocessing

ata are key ingredients for any analytical exercise. Hence, it is
important to thoroughly consider and list all data sources that are
of potential interest before starting the analysis. The rule here is
the more data, the better. However, real life data can be dirty because
of inconsistencies, incompleteness, duplication, and merging problems.
Throughout the analytical modeling steps, various data filtering mechanisms will be applied to clean up and reduce the data to a manageable
and relevant size. Worth mentioning here is the garbage in, garbage
out (GIGO) principle, which essentially states that messy data will yield
messy analytical models. It is of the utmost importance that every data
preprocessing step is carefully justified, carried out, validated, and documented before proceeding with further analysis. Even the slightest
mistake can make the data totally unusable for further analysis. In what
follows, we will elaborate on the most important data preprocessing
steps that should be considered during an analytical modeling exercise.

TYPES OF DATA SOURCES
As previously mentioned, more data is better to start off the analysis.
Data can originate from a variety of different sources, which will be
explored in what follows.
13

▸ A N A LY T I C S IN A BI G DATA WO RL D

Transactions are the first important source of data. Transactional
data consist of structured, low‐level, detailed information capturing
the key characteristics of a customer transaction (e.g., purchase, claim,
cash transfer, credit card payment). This type of data is usually stored
in massive online transaction processing (OLTP) relational databases.
It can also be summarized over longer time horizons by aggregating it
into averages, absolute/relative trends, maximum/minimum values,
and so on.
Unstructured data embedded in text documents (e.g., emails, web
pages, claim forms) or multimedia content can also be interesting to
analyze. However, these sources typically require extensive preprocessing before they can be successfully included in an analytical exercise.
Another important source of data is qualitative, expert‐based
data. An expert is a person with a substantial amount of subject matter expertise within a particular setting (e.g., credit portfolio manager,
brand manager). The expertise stems from both common sense and
business experience, and it is important to elicit expertise as much as
possible before the analytics is run. This will steer the modeling in the
right direction and allow you to interpret the analytical results from
the right perspective. A popular example of applying expert‐based
validation is checking the univariate signs of a regression model. For
example, one would expect a priorii that higher debt has an adverse
impact on credit risk, such that it should have a negative sign in the
final scorecard. If this turns out not to be the case (e.g., due to bad
data quality, multicollinearity), the expert/business user will not be
tempted to use the analytical model at all, since it contradicts prior
expectations.
Nowadays, data poolers are becoming more and more important
in the industry. Popular examples are Dun & Bradstreet, Bureau Van
Dijck, and Thomson Reuters. The core business of these companies
is to gather data in a particular setting (e.g., credit risk, marketing),
build models with it, and sell the output of these models (e.g., scores),
possibly together with the underlying raw data, to interested customers. A popular example of this in the United States is the FICO score,
which is a credit score ranging between 300 and 850 that is provided
by the three most important credit bureaus: Experian, Equifax, and
Transunion. Many financial institutions use these FICO scores either

DATA COLLECTIO N, S AMPLING, AND PREPROCESSING ◂

as their final internal model, or as a benchmark against an internally
developed credit scorecard to better understand the weaknesses of
the latter.
Finally, plenty of publicly available data can be included in the
analytical exercise. A first important example is macroeconomic data
about gross domestic product (GDP), inflation, unemployment, and so
on. By including this type of data in an analytical model, it will become
possible to see how the model varies with the state of the economy.
This is especially relevant in a credit risk setting, where typically all
models need to be thoroughly stress tested. In addition, social media
data from Facebook, Twitter, and others can be an important source
of information. However, one needs to be careful here and make sure
that all data gathering respects both local and international privacy
regulations.

SAMPLING
The aim of sampling is to take a subset of past customer data and use
that to build an analytical model. A first obvious question concerns the
need for sampling. With the availability of high performance computing facilities (e.g., grid/cloud computing), one could also directly analyze the full data set. However, a key requirement for a good sample
is that it should be representative of the future customers on which
the analytical model will be run. Hence, the timing aspect becomes
important because customers of today are more similar to customers
of tomorrow than customers of yesterday. Choosing the optimal time
window for the sample involves a trade‐off between lots of data (and
hence a more robust analytical model) and recent data (which may be
more representative). The sample should also be taken from an average business period to get a picture of the target population that is as
accurate as possible.
It speaks for itself that sampling bias should be avoided as much
as possible. However, this is not always straightforward. Let’s take
the example of credit scoring. Assume one wants to build an application scorecard to score mortgage applications. The future population
then consists of all customers who come to the bank and apply for
a mortgage—the so‐called through‐the‐door (TTD) population. One

▸ A N A LY T I C S IN A BI G DATA WO RL D

Through-the-Door
Rejects
? Bads

Accepts
? Goods

Bads

Goods

Figure 2.1 The Reject Inference Problem in Credit Scoring

then needs a subset of the historical TTD population to build an analytical model. However, in the past, the bank was already applying
a credit policy (either expert based or based on a previous analytical
model). This implies that the historical TTD population has two subsets:
the customers that were accepted with the old policy, and the ones that
were rejected (see Figure 2.1). Obviously, for the latter, we don’t know
the target value since they were never granted the credit. When building a sample, one can then only make use of those that were accepted,
which clearly implies a bias. Procedures for reject inference have been
suggested in the literature to deal with this sampling bias problem.1
Unfortunately, all of these procedures make assumptions and none of
them works perfectly. One of the most popular solutions is bureau‐
based inference, whereby a sample of past customers is given to the
credit bureau to determine their target label (good or bad payer).
When thinking even closer about the target population for credit
scoring, another forgotten subset are the withdrawals. These are
the customers who were offered credit but decided not to take it
(despite the fact that they may have been classified as good by the
old scorecard). To be representative, these customers should also be
included in the development sample. However, to the best of our
knowledge, no procedures for withdrawal inference are typically
applied in the industry.
In stratified sampling, a sample is taken according to predefined
strata. Consider, for example, a churn prediction or fraud detection
context in which data sets are typically very skewed (e.g., 99 percent
nonchurners and 1 percent churners). When stratifying according to
the target churn indicator, the sample will contain exactly the same
percentages of churners and nonchurners as in the original data.

DATA COLLECTION, SAMPLING, AND PREPROCESSING ◂

TYPES OF DATA ELEMENTS
It is important to appropriately consider the different types of data elements at the start of the analysis. The following types of data elements
can be considered:
■

Continuous: These are data elements that are defined on an
interval that can be limited or unlimited. Examples include
income, sales, RFM (recency, frequency, monetary).

■

Categorical
■

Nominal: These are data elements that can only take on a
limited set of values with no meaningful ordering in between.
Examples include marital status, profession, purpose of loan.

■

Ordinal: These are data elements that can only take on a limited set of values with a meaningful ordering in between.
Examples include credit rating; age coded as young, middle
aged, and old.

■

Binary: These are data elements that can only take on two
values. Examples include gender, employment status.

Appropriately distinguishing between these different data elements
is of key importance to start the analysis when importing the data
into an analytics tool. For example, if marital status were to be incorrectly specified as a continuous data element, then the software would
calculate its mean, standard deviation, and so on, which is obviously
meaningless.

VISUAL DATA EXPLORATION AND EXPLORATORY
STATISTICAL ANALYSIS
Visual data exploration is a very important part of getting to know
your data in an “informal” way. It allows you to get some initial
insights into the data, which can then be usefully adopted throughout
the modeling. Different plots/graphs can be useful here. A first popular example is pie charts. A pie chart represents a variable’s distribution as a pie, whereby each section represents the portion of the total
percent taken by each value of the variable. Figure 2.2 represents a pie
chart for a housing variable for which one’s status can be own, rent, or

▸ A N A LY T I C S IN A BI G DATA WO RL D

Goods

Total Population

Own
Rent
For Free

Bads

Own
Rent
For Free

Figure 2.2 Pie Charts for Exploratory Data Analysis

for free (e.g., live with parents). By doing a separate pie chart analysis
for the goods and bads, respectively, one can see that more goods own
their residential property than bads, which can be a very useful starting insight. Bar charts represent the frequency of each of the values
(either absolute or relative) as bars. Other handy visual tools are histograms and scatter plots. A histogram provides an easy way to visualize
the central tendency and to determine the variability or spread of the
data. It also allows you to contrast the observed data with standard
known distributions (e.g., normal distribution). Scatter plots allow you
to visualize one variable against another to see whether there are any
correlation patterns in the data. Also, OLAP‐based multidimensional
data analysis can be usefully adopted to explore patterns in the data.
A next step after visual analysis could be inspecting some basic
statistical measurements, such as averages, standard deviations, minimum, maximum, percentiles, and confidence intervals. One could
calculate these measures separately for each of the target classes

DATA COLLECTIO N, S AMPLING, AND PREPROCESSING ◂

(e.g., good versus bad customer) to see whether there are any interesting patterns present (e.g., whether bad payers usually have a lower
average age than good payers).

MISSING VALUES
Missing values can occur because of various reasons. The information
can be nonapplicable. For example, when modeling time of churn,
this information is only available for the churners and not for the nonchurners because it is not applicable there. The information can also
be undisclosed. For example, a customer decided not to disclose his or
her income because of privacy. Missing data can also originate because
of an error during merging (e.g., typos in name or ID).
Some analytical techniques (e.g., decision trees) can directly deal
with missing values. Other techniques need some additional preprocessing. The following are the most popular schemes to deal with missing values:2
■

Replace (impute). This implies replacing the missing value
with a known value (e.g., consider the example in Table 2.1).
One could impute the missing credit bureau scores with the
average or median of the known values. For marital status, the
mode can then be used. One could also apply regression‐based
imputation whereby a regression model is estimated to model
a target variable (e.g., credit bureau score) based on the other
information available (e.g., age, income). The latter is more
sophisticated, although the added value from an empirical viewpoint (e.g., in terms of model performance) is questionable.

■

Delete. This is the most straightforward option and consists of
deleting observations or variables with lots of missing values. This,
of course, assumes that information is missing at random and has
no meaningful interpretation and/or relationship to the target.

■

Keep. Missing values can be meaningful (e.g., a customer did
not disclose his or her income because he or she is currently
unemployed). Obviously, this is clearly related to the target
(e.g., good/bad risk or churn) and needs to be considered as a
separate category.

▸ A N A LY T I C S IN A BI G DATA WO RL D

Table 2.1

Dealing with Missing Values

Marital
Status

Credit Bureau
Score

Age

Income

1,800

1,200

Single

Nonchurner

1,000

Single

Nonchurner

2,200

Widowed

2,000

Married

Nonchurner

1,200

Single

Nonchurner

1,500

Married

350

Nonchurner

2,100

620

700

Class
Churner

Churner

Single

Churner

Divorced

Nonchurner

As a practical way of working, one can first start with statistically
testing whether missing information is related to the target variable
(using, for example, a chi‐squared test, discussed later). If yes, then we
can adopt the keep strategy and make a special category for it. If not,
one can, depending on the number of observations available, decide to
either delete or impute.

OUTLIER DETECTION AND TREATMENT
Outliers are extreme observations that are very dissimilar to the rest of
the population. Actually, two types of outliers can be considered:
1. Valid observations (e.g., salary of boss is $1 million)
2. Invalid observations (e.g., age is 300 years)
Both are univariate outliers in the sense that they are outlying on
one dimension. However, outliers can be hidden in unidimensional
views of the data. Multivariate outliers are observations that are outlying in multiple dimensions. Figure 2.3 gives an example of two outlying observations considering both the dimensions of income and age.
Two important steps in dealing with outliers are detection and treatment. A first obvious check for outliers is to calculate the minimum
and maximum values for each of the data elements. Various graphical

DATA COLLECTION, SAMPLING, AND PREPROCESSING ◂

Income and Age
4,500
4,000
3,500
3,000
2,500
2,000
1,500
1,000
500
0

Figure 2.3 Multivariate Outliers

tools can be used to detect outliers. Histograms are a first example.
Figure 2.4 presents an example of a distribution for age whereby the
circled areas clearly represent outliers.
Another useful visual mechanism are box plots. A box plot represents three key quartiles of the data: the first quartile (25 percent of
the observations have a lower value), the median (50 percent of the
observations have a lower value), and the third quartile (75 percent
of the observations have a lower value). All three quartiles are represented as a box. The minimum and maximum values are then also

Age

3,500
3,000

Frequency

2,500
2,000
1,500
1,000
500
0

0–5

20–25

25–30

30–35

35–40

40–45

Figure 2.4 Histograms for Outlier Detection

45–50

50–55

55–60

60–65

65–70 150–200

▸ A N A LY T I C S I N A BI G D ATA WO RL D

1.5 * IQR

Min
Figure 2.5

Q1 M

Outliers

Box Plots for Outlier Detection

added unless they are too far away from the edges of the box. Too
far away is then quantified as more than 1.5 * Interquartile Range
(IQR = Q3 − Q1). Figure 2.5 gives an example of a box plot in which
three outliers can be seen.
Another way is to calculate z‐scores, measuring how many standard deviations an observation lies away from the mean, as follows:
xi − μ
σ
where μ represents the average of the variable and σ its standard deviation. An example is given in Table 2.2. Note that by definition, the
z‐scores will have 0 mean and unit standard deviation.
A practical rule of thumb then defines outliers when the absolute
value of the z‐score
z
|z| is bigger than 3. Note that the z‐score relies on
the normal distribution.
The above methods all focus on univariate outliers. Multivariate
outliers can be detected by fitting regression lines and inspecting the
zi =

Table 2.2

Z‐Scores for Outlier Detection

Age

Z‐Score

(30 − 40)/10 = −1

(50 − 40)/10 = +1

(10 − 40)/10 = −3

(40 − 40)/10 = 0

(60 − 40)/10 = +2

(80 − 40)/10 = +4

6
…

…

…
μ = 40
σ = 10

μ=0
σ=1

DATA COLLECTION, SAMPLING, AND PREPROCESSING ◂

μ – 3σ

μ + 3σ

Figure 2.6 Using the Z‐Scores for Truncation

observations with large errors (using, for example, a residual plot).
Alternative methods are clustering or calculating the Mahalanobis distance. Note, however, that although potentially useful, multivariate
outlier detection is typically not considered in many modeling exercises due to the typical marginal impact on model performance.
Some analytical techniques (e.g., decision trees, neural networks, Support Vector Machines (SVMs)) are fairly robust with
respect to outliers. Others (e.g., linear/logistic regression) are more
sensitive to them. Various schemes exist to deal with outliers. It
highly depends on whether the outlier represents a valid or invalid
observation. For invalid observations (e.g., age is 300 years), one
could treat the outlier as a missing value using any of the schemes
discussed in the previous section. For valid observations (e.g.,
income is $1 million), other schemes are needed. A popular scheme
is truncation/capping/winsorizing. One hereby imposes both a
lower and upper limit on a variable and any values below/above
are brought back to these limits. The limits can be calculated using
the z‐scores (see Figure 2.6), or the IQR (which is more robust than
the z‐scores), as follows:
Upper/lower limit = M ± 3s, with M = median and s = IQR/(2 × 0.6745).3
A sigmoid transformation ranging between 0 and 1 can also be
used for capping, as follows:
1
f ( x) =
1 + e− x

▸ A N A LY T I C S IN A BI G DATA WO RL D

In addition, expert‐based limits based on business knowledge and/
or experience can be imposed.

STANDARDIZING DATA
Standardizing data is a data preprocessing activity targeted at scaling
variables to a similar range. Consider, for example, two variables: gender (coded as 0/1) and income (ranging between $0 and $1 million).
When building logistic regression models using both information elements, the coefficient for income might become very small. Hence, it
could make sense to bring them back to a similar scale. The following
standardization procedures could be adopted:
■

Min/max standardization
■

■

Z‐score standardization
■

■

X old − min(X old )
(newmax − newmin) + newmin,
max(X old ) − min(X old )
whereby newmax and newmin are the newly imposed maximum and minimum (e.g., 1 and 0).
X new =

Calculate the z‐scores (see the previous section)

Decimal scaling
■

X old
, with n the
10n
number of digits of the maximum absolute value.
Dividing by a power of 10 as follows: X new =

Again note that standardization is especially useful for regression‐
based approaches, but is not needed for decision trees, for example.

CATEGORIZATION
Categorization (also known as coarse classification, classing, grouping,
binning, etc.) can be done for various reasons. For categorical variables, it is needed to reduce the number of categories. Consider, for
example, the variable “purpose of loan” having 50 different values.
When this variable would be put into a regression model, one would
need 49 dummy variables (50 − 1 because of the collinearity), which
would necessitate the estimation of 49 parameters for only one variable. With categorization, one would create categories of values such

DATA COLLECTION, SAMPLING, AND PREPROCESSING ◂

30
25
20
15
10
5
0
16

Figure 2.7 Default Risk versus Age

that fewer parameters will have to be estimated and a more robust
model is obtained.
For continuous variables, categorization may also be very beneficial. Consider, for example, the age variable and its risk as depicted in
Figure 2.7. Clearly, there is a nonmonotonous relation between risk
and age. If a nonlinear model (e.g., neural network, support vector
machine) were to be used, then the nonlinearity can be perfectly modeled. However, if a regression model were to be used (which is typically more common because of its interpretability), then since it can
only fit a line, it will miss out on the nonmonotonicity. By categorizing
the variable into ranges, part of the nonmonotonicity can be taken
into account in the regression. Hence, categorization of continuous
variables can be useful to model nonlinear effects into linear models.
Various methods can be used to do categorization. Two very basic
methods are equal interval binning and equal frequency binning.
Consider, for example, the income values 1,000, 1,200, 1,300, 2,000,
1,800, and 1,400. Equal interval binning would create two bins with
the same range—Bin 1: 1,000, 1,500 and Bin 2: 1,500, 2,000—whereas
equal frequency binning would create two bins with the same number of observations—Bin 1: 1,000, 1,200, 1,300; Bin 2: 1,400, 1,800,
2,000. However, both methods are quite basic and do not take into
account a target variable (e.g., churn, fraud, credit risk).
Chi‐squared analysis is a more sophisticated way to do coarse classification. Consider the example depicted in Table 2.3 for coarse classifying a residential status variable.

▸ A N A LY T I C S I N A BI G D ATA WO RL D

Table 2.3

Coarse Classifying the Residential Status Variable

Attribute Owner

Rent
Rent
With
Unfurnished Furnished Parents
350

950

Other

No
Answer

Total

9,000

6,000

1,600

Bads

300

400

140

100

1,000

Good:
bad odds

20:1

4:1

2.5:1

9.5:1

1.8:1

1:1

9:1

Source: L. C. Thomas, D. Edelman, and J. N. Crook, Credit Scoring and its Applicationss (Society for Industrial and Applied
Mathematics, Philadelphia, Penn., 2002).

Suppose we want three categories and consider the following
options:
■

Option 1: owner, renters, others

■

Option 2: owner, with parents, others

Both options can now be investigated using chi‐squared analysis.
The purpose is to compare the empirically observed with the independence frequencies. For option 1, the empirically observed frequencies
are depicted in Table 2.4.
The independence frequencies can be calculated as follows. The
number of good owners, given that the odds are the same as in the
whole population, is 6,300/10,000 × 9,000/10,000 × 10,000 = 5,670.
One then obtains Table 2.5.
The more the numbers in both tables differ, the less independence,
hence better dependence and a better coarse classification. Formally,
one can calculate the chi‐squared distance as follows:
χ2 =

(6000 − 5670)2 (300 − 630)2 (1950 − 2241)2 (540 − 249)2
+
+
+
5670
630
2241
249
(1050 − 1089)2 (160 − 121)2
+
+
= 583
1089
121

Table 2.4

Empirical Frequencies Option 1 for Coarse Classifying Residential Status

Attribute

Owner

Renters

Others

Total

Goods

6,000

1,950

1,050

9,000

Bads

300

540

160

1,000

Total

6,300

2,490

1,210

10,000

DATA COLLECTIO N, S AMPLING, AND PREPROCESSING ◂

Table 2.5 Independence Frequencies Option 1 for Coarse Classifying Residential Status

Attribute

Owner

Renters

Others

Total

Goods

5,670

2,241

1,089

9,000

Bads

630

249

121

1,000

Total

6,300

2,490

1,210

10,000

Likewise, for option 2, the calculation becomes:
χ2 =

(6000 − 5670)2 (300 − 630)2 (950 − 945)2 (100 − 105)2
+
+
+
5670
630
945
105
(2050 − 2385)2 (600 − 265)2
+
+
= 662
2385
265

So, based upon the chi‐squared values, option 2 is the better categorization. Note that formally, one needs to compare the value with
a chi‐squared distribution with k − 1 degrees of freedom with k being
the number of values of the characteristic.
Many analytics software tools have built‐in facilities to do categorization using chi‐squared analysis. A very handy and simple approach
(available in Microsoft Excel) is pivot tables. Consider the example
shown in Table 2.6.
One can then construct a pivot table and calculate the odds as
shown in Table 2.7.

Table 2.6

Coarse Classifying the Purpose Variable

Customer ID

Age

Purpose

…

G/B

Car

Cash

Travel

Car

Study

House

Cash

Car

…

▸ A N A LY T I C S IN A BI G DATA WO RL D

Table 2.7 Pivot Table for Coarse Classifying the Purpose Variable

Good
Bad
Odds

Car

Cash

Travel

Study

House

1,000

2,000

3,000

100

5,000

500

100

200

800

1.25

6.25

…

We can then categorize the values based on similar odds. For
example, category 1 (car, study), category 2 (house), and category 3
(cash, travel).

WEIGHTS OF EVIDENCE CODING
Categorization reduces the number of categories for categorical variables. For continuous variables, categorization will introduce new
variables. Consider a regression model with age (4 categories, so 3
parameters) and purpose (5 categories, so 4 parameters) characteristics. The model then looks as follows:
Y = β0 + β1Age1 + β2Age2 + β3Age3 + β4 Purp1
+ β5Purp2 + β6Purp3 + β7Purp4
Despite having only two characteristics, the model still needs 8
parameters to be estimated. It would be handy to have a monotonic
transformation f(.) such that our model could be rewritten as follows:
Y = β0 + β1 f (Age1 , Age2 , Age3) + β2 f (Purp1 , Purp2 , Purp3 , Purp4 )
The transformation should have a monotonically increasing or
decreasing relationship with Y. Weights‐of‐evidence coding is one
example of a transformation that can be used for this purpose. This is
illustrated in Table 2.8.
The WOE is calculated as: ln(Distr. Good/Distr. Bad). Because of
the logarithmic transformation, a positive (negative) WOE means
Distr. Good > (<) Distr. Bad. The WOE transformation thus implements a transformation monotonically related to the target variable.
The model can then be reformulated as follows:
Y = β0 + β1WOEage + β2WOEpurpose

DATA COLLECTIO N, S AMPLING, AND PREPROCESSING ◂

Table 2.8

Calculating Weights of Evidence (WOE)

Count

Distr.
Count

Goods

Distr.
Good

Bads

Distr.
Bad

WOE

2.50%

2.33%

4.12%

−57.28%

18–22

200

10.00%

152

8.42%

24.74%

−107.83%

23–26

300

15.00%

246

13.62%

27.84%

−71.47%

27–29

450

22.50%

405

22.43%

23.20%

−3.38%

30–35

500

25.00%

475

26.30%

12.89%

71.34%

35–44

350

17.50%

339

18.77%

5.67%

119.71%

150

7.50%

147

8.14%

1.55%

166.08%

Age

44+

2,000

1,806

194

This gives a more concise model than the model with which we
started this section. However, note that the interpretability of the
model becomes somewhat less straightforward when WOE variables
are being used.

VARIABLE SELECTION
Many analytical modeling exercises start with tons of variables, of
which typically only a few actually contribute to the prediction of
the target variable. For example, the average application/behavioral
scorecard in credit scoring has somewhere between 10 and 15 variables. The key question is how to find these variables. Filters are a
very handy variable selection mechanism. They work by measuring
univariate correlations between each variable and the target. As such,
they allow for a quick screening of which variables should be retained
for further analysis. Various filter measures have been suggested in the
literature. One can categorize them as depicted in Table 2.9.
The Pearson correlation ρP is calculated as follows:

∑ i =1(X i − X )(Yi − Y )
n
n
∑ i =1(X i − X )2 ∑ i =1(Yi − Y )2
n

ρP =

varies between −1 and +1. To apply it as a filter, one could select all
variables for which the Pearson correlation is significantly different

▸ A N A LY T I C S IN A BI G DATA WO RL D

Table 2.9

Filters for Variable Selection

Continuous Target
(e.g., CLV, LGD)

Categorical Target (e.g.,
churn, fraud, credit risk)

Continuous variable

Pearson correlation

Fisher score

Categorical variable

Fisher score/ANOVA

Information value (IV)
Cramer’s V
Gain/entropy

from 0 (according to the p‐value), or, for example, the ones where
|ρP| > 0.50.
The Fisher score can be calculated as follows:
XG − X B
sG2 + sB2

where X G (X B ) represents the average value of the variable for the
Goods (Bads) and sG2 (sB2 ) the corresponding variances. High values of
the Fisher score indicate a predictive variable. To apply it as a filter,
one could, for example, keep the top 10 percent. Note that the Fisher
score may generalize to a well‐known analysis of variance (ANOVA) in
case a variable has multiple categories.
The information value (IV) filter is based on weights of evidence
and is calculated as follows:
k

IV = ∑(Dist Goodi − Dist Badi )* WOEi
i =1

k represents the number of categories of the variable. For the
example discussed in Table 2.8, the calculation becomes as depicted in
Table 2.10.
The following rules of thumb apply for the information value:
■

< 0.02: unpredictive

■

0.02–0.1: weak predictive

■

0.1–0.3: medium predictive

■

> 0.3: strong predictive

Note that the information value assumes that the variable has
been categorized. It can actually also be used to adjust/steer the categorization so as to optimize the IV. Many software tools will provide

DATA COLLECTION, SAMPLING, AND PREPROCESSING ◂

Table 2.10

Age

Calculating the Information Value Filter Measure

Count

Missing 50

Distr.
Count

Goods

2.50%

Distr.
Good

Bads

Distr.
Bad

WOE

2.33%

4.12%

−57.28%

0.0103

18–22

200

10.00%

152

8.42%

24.74% −107.83%

0.1760

23–26

300

15.00%

246

13.62%

27.84%

−71.47%

0.1016

27–29

450

22.50%

405

22.43%

23.20%

−3.38%

0.0003

30–35

500

25.00%

475

26.30%

12.89%

71.34%

0.0957

35–44

350

17.50%

339

18.77%

5.67%

119.71%

0.1568

44+

150

7.50%

147

8.14%

1.55%

166.08%

0.1095

Information Value

0.6502

interactive support to do this, whereby the modeler can adjust the
categories and gauge the impact on the IV. To apply it as a filter, one
can calculate the information value of all (categorical) variables and
only keep those for which the IV > 0.1 or, for example, the top 10%.
Another filter measure based upon chi‐squared analysis is Cramer’s
V. Consider the contingency table depicted in Table 2.11 for marital
status versus good/bad.
Similar to the example discussed in the section on categorization,
the chi‐squared value for independence can then be calculated as
follows:
χ2 =

(500 − 480)2 (100 − 120)2 (300 − 320)2 (100 − 80)2
+
+
+
= 10.41
480
120
320
80

k − 1 degrees of freedom, with k being the number of classes of the characteristic. The
Cramer’s V measure can then be calculated as follows:
Cramer ′s V =
Table 2.11

χ2
= 0.10,
n

Contingency Table for Marital Status versus Good/Bad Customer

Good

Bad

Total

Married

500

100

600

Not Married

300

100

400

Total

800

200

1,000

▸ A N A LY T I C S IN A BI G DATA WO RL D

with n being the number of observations in the data set. Cramer’s V
is always bounded between 0 and 1 and higher values indicate better predictive power. As a rule of thumb, a cutoff of 0.1 is commonly
adopted. One can then again select all variables where Cramer’s V is
bigger than 0.1, or consider the top 10 percent. Note that the information value and Cramer’s V typically consider the same characteristics
as most important.
Filters are very handy because they allow you to reduce the number of dimensions of the data set early in the analysis in a quick way.
Their main drawback is that they work univariately and typically do
not consider, for example, correlation between the dimensions individually. Hence, a follow-up input selection step during the modeling
phase will be necessary to further refine the characteristics. Also worth
mentioning here is that other criteria may play a role in selecting variables. For example, from a regulatory compliance viewpoint, some
variables may not be used in analytical models (e.g., the U.S. Equal
Credit Opportunities Act states that one cannot discriminate credit
based on age, gender, marital status, ethnic origin, religion, and so on,
so these variables should be left out of the analysis as soon as possible).
Note that different regulations may apply in different geographical
regions and hence should be checked. Also, operational issues could
be considered (e.g., trend variables could be very predictive but may
require too much time to be computed in a real‐time online scoring
environment).

SEGMENTATION
Sometimes the data is segmented before the analytical modeling
starts. A first reason for this could be strategic (e.g., banks might
want to adopt special strategies to specific segments of customers).
It could also be motivated from an operational viewpoint (e.g., new
customers must have separate models because the characteristics
in the standard model do not make sense operationally for them).
Segmentation could also be needed to take into account significant
variable interactions (e.g., if one variable strongly interacts with a
number of others, it might be sensible to segment according to this
variable).

DATA COLLECTION, SAMPLING, AND PREPROCESSING ◂

The segmentation can be conducted using the experience and
knowledge from a business expert, or it could be based on statistical
analysis using, for example, decision trees (see Chapter 3), k‐means, or
self‐organizing maps (see Chapter 4).
Segmentation is a very useful preprocessing activity because one
can now estimate different analytical models each tailored to a specific
segment. However, one needs to be careful with it because by segmenting, the number of analytical models to estimate will increase,
which will obviously also increase the production, monitoring, and
maintenance costs.

NOTES
1. J. Banasik, J. N. Crook, and L. C. Thomas, “Sample Selection Bias in Credit Scoring Models” in Proceedings of the Seventh Conference on Credit Scoring and Credit Control
(Edinburgh University, 2001).
2. R. J. A. Little and D. B. Rubin, Statistical Analysis with Missing Data (Wiley-Interscience, Hoboken, New Jersey, 2002).
3. T. Van Gestel and B. Baesens, Credit Risk Management: Basic Concepts: Financial Risk
Components, Rating Analysis, Models, Economic and Regulatory Capital, Oxford University
Press, Oxford, England, ISBN 978-0-19-954511-7, 2009.

C H A P T E R

Predictive
Analytics

n predictive analytics, the aim is to build an analytical model predicting a target measure of interest.1 The target is then typically
used to steer the learning process during an optimization procedure.
Two types of predictive analytics can be distinguished: regression and
classification. In regression, the target variable is continuous. Popular examples are predicting stock prices, loss given default (LGD), and
customer lifetime value (CLV). In classification, the target is categorical. It can be binary (e.g., fraud, churn, credit risk) or multiclass (e.g.,
predicting credit ratings). Different types of predictive analytics techniques have been suggested in the literature. In what follows, we will
discuss a selection of techniques with a particular focus on the practitioner’s perspective.

TARGET DEFINITION
Because the target variable plays an important role in the learning
process, it is of key importance that it is appropriately defined. In what
follows, we will give some examples.
In a customer attrition setting, churn can be defined in various ways. Active churn implies that the customer stops the relationship with the firm. In a contractual setting (e.g., postpaid telco),

▸ A N A LY T I C S IN A BI G DATA WO RL D

this can be easily detected when the customer cancels the contract.
In a noncontractual setting (e.g., supermarket), this is less obvious
and needs to be operationalized in a specific way. For example, a
customer churns if he or she has not purchased any products during
the previous three months. Passive churn occurs when a customer
decreases the intensity of the relationship with the firm, for example, by decreasing product or service usage. Forced churn implies
that the company stops the relationship with the customer because
he or she has been engaged in fraudulent activities. Expected churn
occurs when the customer no longer needs the product or service
(e.g., baby products).
In credit scoring, a defaulter can be defined in various ways. For
example, according to the Basel II/Basel III regulation, a defaulter
is defined as someone who is 90 days in payment arrears. In the
United States, this has been changed to 180 days for mortgages and
qualifying revolving exposures, and 120 days for other retail exposures. Other countries (e.g., the United Kingdom) have made similar
adjustments.
In fraud detection, the target fraud indicator is usually hard to
determine because one can never be fully sure that a certain transaction (e.g., credit card) or claim (e.g., insurance) is fraudulent. Typically,
the decision is then made based on a legal judgment or a high suspicion by a business expert.2
In response modeling, the response target can be defined in various ways. Gross response refers to the customers who purchase after
having received the marketing message. However, it is more interesting to define the target as the net response, being the customers who
purchase because of having received the marketing message, the so‐
called swingers.
Customer lifetime value (CLV) is a continuous target variable and
is usually defined as follows:3
(Rt − C t )st
t
i =1 (1 + d)
n

CLV = ∑

where n represents the time horizon considered (typically two to three
years), Rt the revenue at time t (both direct and indirect), Ct the costs
incurred at time t (both direct and indirect), st the survival probability

P R E D I C T I V E A N A LY T I C S ◂

Table 3.1

Example CLV Calculation

Revenue in
Month t (Rt)

Cost in Month
t ( C t)

Survival
Probability in
Month t (st)

( Rt − C t) *
st / (1 + d )t

150

0.94

135.22

100

0.92

82.80

120

0.88

101.20

100

0.84

84.00

130

0.82

98.40

140

0.74

99.90

0.7

45.50

100

0.68

61.20

120

0.66

72.60

0.6

42.00

100

0.55

55.00

130

0.5

60.00

CLV

937.82

Month t
1

Yearly WACC

10%

Monthly WACC

at time t (see Chapter 5), and d the discounting factor (typically the
weighted average cost of capital [WACC]). Defining all these parameters is by no means a trivial exercise and should be done in close
collaboration with the business expert. Table 3.1 gives an example of
calculating CLV.
Loss given default (LGD) is an important credit risk parameter in a
Basel II/Basel III setting.4 It represents the percentage of the exposure
likely to be lost upon default. Again, when defining it, one needs to
decide on the time horizon (typically two to three years), what costs
to include (both direct and indirect), and what discount factor to adopt
(typically the contract rate).
Before starting the analytical step, it is really important to check
the robustness and stability of the target definition. In credit scoring,
one commonly adopts roll rate analysis for this purpose as illustrated
in Figure 3.1. The purpose here is to visualize how customers move
from one delinquency state to another during a specific time frame. It

▸ A N A LY T I C S IN A BI G DATA WO RL D

Worst—Previous 12 Months

Roll Rate

90+
60 day
30 day
Curr/x day
0%

20%

40%

60%

80%

100%

Worst—Next 12 Months
Curr/x day

30 day

60 day

90+

Figure 3.1 Roll Rate Analysis
Source: N. Siddiqi, Credit Risk Scorecards: Developing and Implementing Intelligent Credit Scoring
(Hoboken, NJ: John Wiley & Sons, 2005).

can be easily seen from the plot that once the customer has reached
90 or more days of payment arrears, he or she is unlikely to recover.

LINEAR REGRESSION
Linear regression is a baseline modeling technique to model a continuous target variable. For example, in a CLV modeling context, a linear
regression model can be defined to model CLV in terms of the RFM
(recency, frequency, monetary value) predictors as follows:
CLV = β0 + β1R + β2F + β3 M
The β parameters are then typically estimated using ordinary least
squares (OLS) to minimize the sum of squared errors. As part of the
estimation, one then also obtains standard errors, p‐values indicating
variable importance (remember important variables get low p‐values),
and confidence intervals. A key advantage of linear regression is that it
is simple and usually works very well.
Note that more sophisticated variants have been suggested in the
literature (e.g., ridge regression, lasso regression, time series models [ARIMA, VAR, GARCH], multivariate adaptive regression splines
[MARS]).

P R E D I C T I V E A N A LY T I C S ◂

LOGISTIC REGRESSION
Consider a classification data set for response modeling as depicted in
Table 3.2.
When modeling the response using linear regression, one gets:
Y = β0 + β1 Age + β2 Income + β3 Gender
When estimating this using OLS, two key problems arise:
1. The errors/target are not normally distributed but follow a
Bernoulli distribution.
2. There is no guarantee that the target is between 0 and 1, which
would be handy because it can then be interpreted as a probability.
Consider now the following bounding function:
f (z ) =

1
1 + e− z

which can be seen in Figure 3.2.
For every possible value of z, the outcome is always between
0 and 1. Hence, by combining the linear regression with the bounding
function, we get the following logistic regression model:
P(response = yes | age , income , gender) =

1
1+ e

−(β0 +β1age+β2income+β3gender)

The outcome of the above model is always bounded between 0
and 1, no matter what values of age, income, and gender are being
used, and can as such be interpreted as a probability.

Table 3.2 Example Classification Data Set

Customer

Age

Income

Gender

John

1,200

Sarah

800

Sophie

David
Peter

...

Response

Yes

2,200

Yes

2,000

1,800

Yes

▸ A N A LY T I C S IN A BI G DATA WO RL D

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
–7

Figure 3.2

–5

–3

–1

Bounding Function for Logistic Regression

The general formulation of the logistic regression model then
becomes:
P(Y = 1| X 1 ,… , X n ) =

1
1+ e

−(β0 +β1X1 + +β N X N )

or, alternately,
P(Y = 0| X 1 ,… , X N ) = 1 − P(Y = 1| X 1 ,… , X N )
1
1
= 1−
=
1 + e −(β0 +β1X1+ +βN X N ) 1 + e(β0 +β1X1+ +βN X N
Hence, both P(Y = 1| X 1 ,…, X N ) and P(Y = 0| X 1 ,… , X N ) are bounded
between 0 and 1.
Reformulating in terms of the odds, the model becomes:
P(Y = 1| X 1 ,…, X N )
= e(β0 +β1X1+ +βN X N )
P(Y = 0| X 1 ,…, X N )
or, in terms of log odds (logit),
⎛ P(Y = 1| X 1 ,…, X N ) ⎞
= β0 + β1X 1 + + β N X N
ln ⎜
⎝ P(Y = 0| X 1 ,…, X N )⎟⎠

P R E D I C T I V E A N A LY T I C S ◂

Income

G
B

B G

B
B
B

B
G

G
B

BB
B

G
G
G
G
B B

G
G
G

G
G G
G

B
G

G G

G
G

G
G
G

Age
Figure 3.3

Decision Boundary of Logistic Regression

The βi parameters of a logistic regression model are then estimated
by optimizing a maximum likelihood function. Just as with linear
regression, the optimization comes with standard errors, p‐values for
variable screening and confidence intervals.
Since logistic regression is linear in the log odds (logit), it basically
estimates a linear decision boundary to separate both classes. This is
illustrated in Figure 3.3.
To interpret a logistic regression model, one can calculate the odds
ratio. Suppose variable Xi increases with one unit with all other variables being kept constant (ceteris paribus), then the new logit becomes
the old logit with βi added. Likewise, the new odds become the old
odds multiplied by eβi. The latter represents the odds ratio, that is, the
multiplicative increase in the odds when Xi increases by 1 (ceteris paribus). Hence,
■

βi > 0 implies eβi > 1 and the odds and probability increase with
Xi

■

βi < 0 implies eβi < 1 and the odds and probability decrease
with Xi

Another way of interpreting a logistic regression model is by calculating the doubling amount. This represents the amount of change
required for doubling the primary outcome odds. It can be easily
seen that for a particular variable Xi, the doubling amount equals
log(2)/βi.

▸ A N A LY T I C S IN A BI G DATA WO RL D

Note that next to the f(z) transformation discussed above, other
transformations also have been suggested in the literature. Popular
examples are the probit and cloglog transformation as follows:
f (z ) =

1
2π

∫

f (z ) = 1 − e − e

−∞

−t2
2

The probit transformation was used in Moody’s RiskCalc tool for
predicting probability of default for firms.5 Note, however, that empirical evidence suggests that all three transformations typically perform
equally well.

DECISION TREES
Decision trees are recursive partitioning algorithms (RPAs) that come
up with a tree-like structure representing patterns in an underlying
data set.6 Figure 3.4 provides an example of a decision tree.
The top node is the root node specifying a testing condition
of which the outcome corresponds to a branch leading up to an
internal node. The terminal nodes of the tree assign the classifications and are also referred to as the leaf nodes. Many algorithms
have been suggested to construct decision trees. Amongst the most
popular are: C4.5 (See5),7 CART,8 and CHAID.9 These algorithms
differ in their way of answering the key decisions to build a tree,
which are:
■

Splitting decision: Which variable to split at what value (e.g.,
age < 30 or not, income < 1,000 or not; marital status = married
or not)

■

Stopping decision: When to stop growing a tree?

■

Assignment decision: What class (e.g., good or bad customer) to
assign to a leaf node?

Usually, the assignment decision is the most straightforward to
make since one typically looks at the majority class within the leaf
node to make the decision. The other two decisions to be made are less
straightforward and are elaborated on in what follows.

P R E D I C T I V E A N A LY T I C S ◂

Income > $50,000
No

Yes

Employed

Age < 40

Respond

Yes

Not Respond

Respond

No
Not Respond

Figure 3.4 Example Decision Tree

In order to answer the splitting decision, one needs to define the
concept of impurity or chaos. Consider, for example, the three data sets
of Figure 3.5, each of which contains good (unfilled circles) and bad
(filled circles) customers. Minimal impurity occurs when all customers
are either good or bad. Maximal impurity occurs when one has the same
number of good and bad customers (i.e., the data set in the middle).
Decision trees will now aim at minimizing the impurity in the data.
In order to do so appropriately, one needs a measure to quantify impurity. Various measures have been introduced in the literature, and the
most popular are:
■

Entropy: E(S) = −pGlog2(pG)−pBlog2(pB) (C4.5/See5)

■

Gini: Gini(S) = 2pGpB (CART)

■

Chi‐squared analysis (CHAID)

with pG (pB) being the proportions of good and bad, respectively. Both
measures are depicted in Figure 3.6, where it can be clearly seen that
the entropy (Gini) is minimal when all customers are either good or bad,
and maximal in the case of the same number of good and bad customers.
Minimal Impurity

Figure 3.5

Maximal Impurity

Example Data Sets for Calculating Impurity

Minimal Impurity

▸ A N A LY T I C S IN A BI G DATA WO RL D

44
1
0.9
0.8
0.7
0.6

Entropy
Gini

0.5
0.4
0.3
0.2
0.1
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Figure 3.6 Entropy versus Gini

In answering the splitting decision, various candidate splits will
now be evaluated in terms of their decrease in impurity. Consider, for
example, a split on age as depicted in Figure 3.7.
The original data set had maximum entropy. The entropy calculations become:
■

Entropy top node = −1/2 × log2(1/2) – 1/2 × log2(1/2) = 1

■

Entropy left node = −1/3 × log2(1/3) – 2/3 × log2(2/3) = 0.91

■

Entropy right node = −1 × log2(1) – 0 × log2(0) = 0

400 400
Age < 30

200 400
B
G
Figure 3.7

Calculating the Entropy for Age Split

Age ≥ 30

200 0
G B

P R E D I C T I V E A N A LY T I C S ◂

The weighted decrease in entropy, also known as the gain, can
then be calculated as follows:
Gain = 1 − (600/800) × 0.91 − (200/800) × 0 = 0.32

Misclassification error

It speaks for itself that a larger gain is to be preferred. The decision
tree algorithm will now consider different candidate splits for its root
node and adopt a greedy strategy by picking the one with the biggest
gain. Once the root node has been decided on, the procedure continues in a recursive way to continue tree growing.
The third decision relates to the stopping criterion. Obviously, if the
tree continues to split, it will become very detailed with leaf nodes containing only a few observations. In other words, the tree will start to fit
the specificities or noise in the data, which is also referred to as overfitting. In order to avoid this, the data will be split into a training sample
and a validation sample. The training sample will be used to make the
splitting decision. The validation sample is an independent sample, set
aside to monitor the misclassification error (or any other performance
metric). One then typically observes a pattern as depicted in Figure 3.8.
The error on the training sample keeps on decreasing as the splits
become more and more specific toward it. On the validation sample,
the error will initially decrease, but at some point it will increase back
again since the splits become too specific for the training sample as the
tree starts to memorize it. Where the validation set curve reaches its
minimum, the procedure should be stopped or overfitting will occur.
Note that besides classification error, one might also use accuracy or

Validation set

STOP growing tree!

Minimum
Training set
Number of tree nodes

Figure 3.8

Using a Validation Set to Stop Growing a Decision Tree

▸ A N A LY T I C S IN A BI G DATA WO RL D

profit based measures on the Y‐axis to make the stopping decision.
Also note that, sometimes, simplicity is preferred above accuracy, and
one can select a tree that does not necessarily have minimum validation set error, but a lower number of nodes.
In the example of Figure 3.4, every node had only two branches.
The advantage of this is that the testing condition can be implemented
as a simple yes/no question. Multiway splits allow for more than two
branches and can provide trees that are wider but less deep. In a read
once decision tree, a particular attribute can be used only once in a
certain tree path. Every tree can also be represented as a rule set since
every path from a root node to a leaf node makes up a simple if/then
rule. These rules can then be easily implemented in all kinds of software packages (e.g., Microsoft Excel).
Decision trees essentially model decision boundaries orthogonal to
the axes. This is illustrated in Figure 3.9 for an example decision tree.
Decision trees can also be used for continuous targets. Consider
the example in Figure 3.10 of a regression tree for predicting LGD.
Other criteria need now be used to make the splitting decision
because the impurity will need to be measured in another way. One
way to measure impurity in a node is by calculating the mean squared
error (MSE) as follows:
1 n
∑(Yi − Y )2 ,
n i =1
where n represents the number of observations in a leave node, Yi the
value of observation i, and Y , the average of all values in the leaf node.

Income

1,200

G
G

B
G

G
G

G
B
B

G
G
G
G

G
G

Income

G
G
G

G
G

Age

B
B

Figure 3.9

G
G

Age

Decision Boundary of a Decision Tree

1,200

P R E D I C T I V E A N A LY T I C S ◂

Loan Collateral
None

Real Estate
Cash
Known Client

Geographic Region
United
States

LGD = 30%

LGD = 18%

Yes

LGD = 42%
Figure 3.10

LGD = 72%

LGD = 55%

Example Regression Tree for Predicting LGD

Another way is by conducting a simple analysis of variance
(ANOVA) test and calculate an F‐statistic as follows:
F=

SSbetween /(B − 1)
∼ Fn− B ,B −1 ,
SSwithin /(n − B)

whereby
B

SSbetween = ∑ nb(Yb − Y )2
b =1

SSwithin = ∑ ∑(Ybi − Yb )2
b =1 i =1

with B being the number of branches of the split, nb the number of
observations in branch b, Yb the average in branch b,Ybi the value of
observation i in branch b, and Y the overall average. Good splits will
then result in a high F value, or low corresponding p‐value.
The stopping decision can be made in a similar way as for classification trees, but using a regression‐based performance measure (e.g.,
mean squared error, mean absolute deviation, R‐squared) on the Y‐
Y
axis. The assignment decision can be made by assigning the mean (or
median) to each leaf node. Note also that confidence intervals may be
computed for each of the leaf nodes.
Decision trees can be used for various purposes in analytics. First,
they can be used for input selection because attributes that occur at the
top of the tree are more predictive of the target. One could also simply calculate the gain of a characteristic to gauge its predictive power.

▸ A N A LY T I C S IN A BI G DATA WO RL D

Next, they can also be used for initial segmentation. One then typically
builds a tree of two or three levels deep as the segmentation scheme
and then uses second stage logistic regression models for further
refinement. Finally, decision trees can also be used as the final analytical model to be used directly into production. A key advantage here is
that the decision tree gives a white box model with a clear explanation
behind how it reaches its classifications. Many software tools will also
allow you to grow trees interactively by providing a splitting option at
each level of the tree (e.g., a top five, or more, of splits amongst which
the modeler can choose). This allows us to choose splits not only based
upon impurity reduction, but also on the interpretability and/or computational complexity of the split criterion.

NEURAL NETWORKS
A first perspective on the origin of neural networks states that they
are mathematical representations inspired by the functioning of the
human brain. Another more realistic perspective sees neural networks
as generalizations of existing statistical models. Let’s take logistic
regression as an example:
P(Y = 1| X 1 ,… , X N ) =

1
1+ e

−(β0 +β1X1 + +β N X N )

This model can be seen in Figure 3.11.
The processing element or neuron in the middle basically performs two operations: it takes the inputs and multiplies them with the
weights (including the intercept term β0, which is called the bias term
X1

β1

X2 β2

P (Y | X 1 ,..., XN ) =

...

β N–1
XN–1

1
1 + e–(β0 +β1X1 + ... + βNXN)

β0

βN

XN
Figure 3.11

Neural Network Representation of Logistic Regression

P R E D I C T I V E A N A LY T I C S ◂

h1
W11

hj = f(Σ xiwij + bj)
i=1

x1
h2

W23

z = Σ vjhj + b4
j=1

Figure 3.12

A Multilayer Perceptron (MLP) Neural Network

in neural networks) and then puts this into a nonlinear transformation function similar to the one we discussed in the section on logistic
regression. So, logistic regression is a neural network with one neuron.
Similarly, we could visualize linear regression as a one neuron neural
network with the identity transformation f(z) = z. We can now generalize the above picture to a multilayer perceptron (MLP) neural network by adding more layers and neurons as shown in Figure 3.12.10
The example in Figure 3.12 is an MLP with one input layer, one
hidden layer, and one output layer. The hidden layer has a nonlinear
transformation function f( ) and the output layer a linear transformation function. The most popular transformation functions (also called
squashing, activation functions) are:
1
■ Logistic, f (z ) =
, ranging between 0 and 1
1 + e−z
ez − e−z
■ Hyperbolic tangent, f (z ) = z
, ranging between –1 and +1
e + e−z
■ Linear, f (z ) = z , ranging between −∞ and +∞

For classification (e.g., churn, response, fraud), it is common practice to adopt a logistic transformation in the output layer, since the
outputs can then be interpreted as probabilities.11 For regression targets (e.g., CLV, LGD), one could use any of the transformation functions listed above. Typically, one will use hyperbolic tangent activation
functions in the hidden layer.
In terms of hidden layers, theoretical works have shown that
neural networks with one hidden layer are universal approximators,

▸ A N A LY T I C S I N A BI G D ATA WO RL D

capable of approximating any function to any desired degree of accuracy on a compact interval.12 Only for discontinuous functions (e.g., a
saw tooth pattern), it could make sense to try out more hidden layers,
although these patterns rarely occur in real‐life data sets.
For simple statistical models (e.g., linear regression), there exists a
closed‐form mathematical formula for the optimal parameter values.
However, for neural networks, the optimization is a lot more complex and the weights sitting on the connections need to be estimated
using an iterative algorithm. The algorithm then optimizes a cost function, which may be similar to linear regression (mean squared error)
or logistic regression (maximum likelihood based). The procedure
typically starts from a set of random weights that are then iteratively
adjusted to the patterns in the data using an optimization algorithm.
Popular optimization algorithms here are backpropagation learning,
conjugate gradient, and Levenberg‐Marquardt.13 A key issue to note
here is the curvature of the objective function, which is not convex
and may be multimodal as illustrated in Figure 3.13. The error function can thus have multiple local minima but typically only one global
minimum. Hence, if the starting weights are chosen in a suboptimal
way, one may get stuck in a local minimum. One way to deal with this
is to try out different starting weights, start the optimization procedure
for a few steps, and then continue with the best intermediate solution.
The optimization procedure then continues until the error function
shows no further progress, the weights stop changing substantially, or
after a fixed number of optimization steps (also called epochs).
E

Local minimum!
Global minimum!

w
Figure 3.13 Local versus Global Minima

P R E D I C T I V E A N A LY T I C S ◂

Although multiple output neurons could be used (predicting, for
example, churn and CLV simultaneously), it is highly advised to use
only one. The hidden neurons, however, should be carefully tuned
and depend on the nonlinearity in the data. More complex, nonlinear
patterns will require more hidden neurons. Although various procedures (e.g., cascade correlation, genetic algorithms, Bayesian methods)
have been suggested in the scientific literature to do this, the most
straightforward yet efficient procedure is as follows:14
■

Split the data into a training, validation, and test set.

■

Vary the number of hidden neurons from 1 to 10 in steps of 1
or more.

■

Train a neural network on the training set and measure the performance on the validation set (may be train multiple neural
networks to deal with the local minimum issue).

■

Choose the number of hidden neurons with optimal validation
set performance.

■

Measure the performance on the independent test set.

Neural networks can model very complex patterns and decision
boundaries in the data and, as such, are very powerful. In fact, they
are so powerful that they can even model the noise in the training
data, which is something that definitely should be avoided. One way
to avoid this overfitting is by using a validation set in a similar way as
with decision trees. This is illustrated in Figure 3.14. The training set
is used here to estimate the weights and the validation set is again an
independent data set used to decide when to stop training. Another
scheme to prevent a neural network from overfitting is weight regularization, whereby the idea is to keep the weights small in absolute
Error
Validation set

STOP training!

Minimum
Training set
Training steps
Figure 3.14

Using a Validation Set for Stopping Neural Network Training

▸ A N A LY T I C S I N A BI G D ATA WO RL D

sense because otherwise they may be fitting the noise in the data. This
is then implemented by adding a weight size term (e.g., Euclidean
norm) to the objective function of the neural network.15
Although neural networks have their merits in terms of modeling
power, they are commonly described as black box techniques because
they relate the inputs to the outputs in a mathematically complex, nontransparent, and opaque way. In application areas where interpretability may not be required (e.g., fraud detection, response modeling), they
can be very successfully adopted as high‐performance analytical tools.
However, in application areas where explanation is important
(e.g., credit risk, medical diagnosis), one needs to be careful with neural networks because insight and comprehensibility in the underlying
patterns is crucial.16 Two ways to open up the neural network black
box are rule extraction and two‐stage models.
The purpose of rule extraction is to extract if/then classification
rules mimicking the behavior of the neural network.17 Two important approaches here are decompositional and pedagogical techniques.
Decompositional rule extraction approaches decompose the network’s
internal workings by inspecting weights and/or activation values. A
typical five‐step approach here could be:18
1. Train a neural network and prune it as much as possible in
terms of connections.
2. Categorize the hidden unit activation values using clustering.
3. Extract rules that describe the network outputs in terms of the
categorized hidden unit activation values.
4. Extract rules that describe the categorized hidden unit activation values in terms of the network inputs.
5. Merge the rules obtained in steps 3 and 4 to directly relate the
inputs to the outputs.
This is illustrated in Figure 3.15.
Note that steps 3 and 4 can be done simultaneously by building
a decision tree relating the network outputs to the hidden unit activation values. Figure 3.16 gives an example of applying a decompositional neural network rule extraction approach in a credit scoring
setting.

1,500

Will

Dan

Bob

1,500

Will

Dan

Bob

h1
1.22
–2.34 1

0.16 3

0.82 2
2

Yes

h1 h2 h3 Response

Step 3: Categorize hidden unit activations.

Step 2: Build a neural network
(e.g, 3 hidden neurons).

Figure 3.15

Decompositional Approach for Neural Network Rule Extraction

If age < 28 and income < 1,000 and gender = F then response = No
If age > 34 and income > 1,500 then response = Yes

Step 6: Merge both rule sets

Step 5: Extract rules relating categorized
hidden units to inputs.

0.8

–0.18

h3
0.66 1

Step 1: Start from original data.

If age < 28 and income < 1,000, then h1 = 1
If gender = F, then h2 = 3
If age > 34 and income > 1,500, then h2 = 2

–0.1

h2
2.34

Yes

Step 4: Extract rules relating network outputs
to categorized hidden units.

2.1

0.78

–1.20

If h1 = 1 and h2 = 3, then response = No
If h2 = 2, then response = Yes

2,400

1,200

1,000

Emma

Customer Age Income Gender

2,400

1,200

1,000

Emma

Income Gender … Response

Age

Customer

Figure 3.16

–0.289

0.137

–0.162

–0.081

0.278

0.380

0.611

–0.453

0.457

Default class: applicant = good

If savings account ≤ 12.40 Euro and economical sector = sector C, then
applicant = bad

If purpose = secondhand car and income > 719 Euro and owns property = no and
savings account ≤ 12.40 Euro and years client ≤ 3, then applicant = bad

If purpose = cash provisioning and income > 719 and owns property = no and savings
account ≤ 12.40 Euro and years client ≤ 3, then applicant = bad

If term > 12 months and purpose = cash provisioning and owns property = no and
savings account ≤ 12.40 Euro and years client ≤ 3, then applicant = bad

If term > 12 months and purpose = cash provisioning and savings account ≤ 12.40
Euro and years client ≤ 3, then applicant = bad

Applicant = bad

Applicant = good

Example of Decompositional Neural Network Rule Extraction

Economical sector = sector C

Years client > 3 years

Property = No

Income > 719 Euro

Savings account > 12.40 Euro

–0.102

–0.287

Purpose = cash provisioning

Purpose = second hand car

–0.202

Term > 12 Months

P R E D I C T I V E A N A LY T I C S ◂

Pedagogical rule extraction techniques consider the neural network as a black box and use the neural network predictions as input
to a white box analytical technique such as decision trees.19 This is
illustrated in Figure 3.17.
In this approach, the learning data set can be further augmented
with artificial data, which is then labeled (e.g., classified or predicted)
by the neural network, so as to further increase the number of observations to make the splitting decisions when building the decision tree.
When using either decompositional or pedagogical rule extraction
approaches, the rule sets should be evaluated in terms of their accuracy,
conciseness (e.g., number of rules, number of conditions per rule), and
fidelity. The latter measures to what extent the extracted rule set perfectly mimics the neural network and is calculated as follows:
Neural Network Classification
Rule set
classification

Good

Bad

Good

Bad

Fidelity = (a + d)/(b + c).
It is also important to always benchmark the extracted rules/trees
with a tree built directly on the original data to see the benefit of going
through the neural network.
Another approach to make neural networks more interpretable
is by using a two‐stage model setup.20 The idea here is to estimate an
easy to understand model first (e.g., linear regression, logistic regression). This will give us the interpretability part. In a second stage, a
neural network is used to predict the errors made by the simple model
using the same set of predictors. Both models are then combined in an
additive way, for example, as follows:
■

Target = linear regression (X1, X2, … XN) + neural network
(X1, X2, … XN)

■

Score = logistic regression (X1, X2, … XN) + neural network
(X1, X2, … XN)

This setup provides an ideal balance between model interpretability (which comes from the first part) and model performance (which
comes from the second part). This is illustrated in Figure 3.18.

Will

Dan

Bob

1,500

Will

Dan

Bob

Figure 3.17

Age < 30

Network prediction Network prediction
response = no
response = no

Yes

Yes
Step 4: Extract rules relating network
predictions to original inputs. Generate
additional data where necessary.

Step 3: Get the network predictions and
add them to the data set.

Step 2: Build a neural network.

Step 1: Start from original data.

Network prediction
response = yes

No
Yes

Yes

Response

Yes

Network
Prediction

Yes

Pedagogical Approach for Rule Extraction

Network prediction
response = yes

Gender = Female
Yes
No

Gender

Gender … Response

Income > 1,500

2,400

1,200

1,000

Income

2,400

1,200

1,500

1,000

Emma

Customer Age

Emma

Customer Age Income

Age

Bob

Customer

Emma

Will

Dan

Bob

2,400

1,200

1,500

1,000

0.88

Yes (=1)

Figure 3.18

Bart

1,000

Two‐Stage Models

0.68

−0.32

0.36

−0.44
0.24

Error

0.18

0.76

0.44

Logistic
Regression
Output

No (=0)

Yes (=1)

No (=0)

Response

Yes

Income Gender …

2,400

1,200

Yes

Final
Output

Dan

1,500

Response

Logistic
Regression
NN Output
Customer Age Income Gender … Output

Will

1,000

Income Gender …

−0.18
0.12

Emma

Logistic
Regression
Customer Age Income Gender … Response
Output
1,000
F
No (=0)
Emma
28
0.44
1,500
M
Yes (=1)
Will
44
0.76
1,200
M
No (=0)
Dan
30
0.18
Bob
58
2,400
M
Yes (=1)
0.88

Age

Customer

Step 5: Score new observations by adding up
logistic regression and NN scores.

Step 4: Build NN predicting errors from
logistic regression model.

Step 3: Calculate errors from logistic
regression model.

Step 2: Build logistic regression model.

Step 1: Start from original data.

▸ A N A LY T I C S I N A BI G D ATA WO RL D

SUPPORT VECTOR MACHINES
Two key shortcomings of neural networks are the fact that the objective
function is nonconvex (and hence may have multiple local minima)
and the effort that is needed to tune the number of hidden neurons.
Support vector machines (SVMs) deal with both of these issues.21
The origins of classification SVMs date back to the early dates of
linear programming.22 Consider the following linear program (LP) for
classification:
min e1 + e2 + + eng + + enb
subject to
w1 x i1 + w 2 x i 2 + + w n x in ≥ c − ei ,1 ≤ i ≤ ng
w1 x i1 + w 2 x i 2 + + w n x in ≤ c + ei , ng + 1 ≤ i ≤ ng + nb
ei ≥ 0
The LP assigns the good customers a score above the cut‐off value c,
and the bad customers a score below c. ng and nb represent the number
of goods and bads, respectively. The error variables ei are needed to be
able to solve the program because perfect separation will typically not be
possible. Linear programming has been very popular in the early days of
credit scoring. One of its benefits is that it is easy to include domain or
business knowledge by adding extra constraints to the model.
A key problem with linear programming is that it can estimate
multiple optimal decision boundaries, as illustrated in Figure 3.19, for
a perfectly linearly separable case.
SVMs add an extra objective to the analysis. Consider, for example, the situation depicted in Figure 3.20. It has two hyperplanes sitting at the edges of both classes and a hyperplane in between, which
will serve as the classification boundary. The perpendicular distance
from the first hyperplane H1 to the origin equals | b−1|/|| w ||, whereby
||w
w|| represents the Euclidean norm of w calculated as || w || = w12 + w 22 .
Likewise, the perpendicular distance from H2 to the origin equals | b +
1|/|| w ||. Hence, the margin between both hyperplanes equals 2/|| w ||.
SVMs will now aim at maximizing this margin to pull both classes as
far apart as possible. Maximizing the margin is similar to minimizing

P R E D I C T I V E A N A LY T I C S ◂

x2
+

+
+

Class 2
+

+
+

x
x

Class 1

x1
Figure 3.19

Multiple Separating Hyperplanes

1 N 2
∑ w i . In case of perfect linear separation, the
2 i =1
SVM classifier then becomes as follows.
Consider a training set: { x k , y k }nk =1 with x k ∈ R N and y k ∈ {−1; +1}
The goods (e.g., class +1) should be above hyperplane H1, and the
bads (e.g., class−1) below hyperplane H2, which gives:
|| w ||, or minimizing

w T x k + b ≥ 1, if yk = +1
w T x k + b ≤ 1, if y k = −1

2/||w||

+
+

Class 2
+

+
+

x
x

H1: wTx + b = + 1
x

Class 1

H0: wTx + b = 0
H2: wTx + b = –1
x1

Figure 3.20

SVM Classifier for the Perfectly Linearly Separable Case

▸ A N A LY T I C S IN A BI G DATA WO RL D

Both can be combined as follows:
yk (w T x k + b) ≥ 1
The optimization problem then becomes:
Minimize

1 N 2
∑ wi
2 i =1

subject to yk (w T x k + b) ≥ 1, k = 1…n
This quadratic programming (QP) problem can now be solved
using Lagrangian optimization.23 It is important to note that the
optimization problem has a quadratic cost function, giving a convex
optimization problem with no local minima and only one global minimum. Training points that lie on one of the hyperplanes H1 or H2 are
called support vectors and are essential to the classification. The classification hyperplane itself is H0 and, for new observations, it needs to be
checked whether they are situated above H0, in which case the prediction is +1 or below (prediction −1). This can be easily accomplished
using the sign operator as follows: y(x) = sign (wTx + b).
The SVM classifier discussed thus far assumed perfect separation is
possible, which will of course rarely be the case for real‐life data sets.
In case of overlapping class distributions (as illustrated in Figure 3.21),
the SVM classifier can be extended with error terms as follows:
Minimize

2/||w||

n
1 N 2
w i + C ∑ ei
∑
2 i =1
i =1

+
+

Class 2
+

+
+

x
+

H1: wTx + b = + 1

+
x

Class 1

H0: wTx + b = 0
H2: wTx + b = –1
x1

Figure 3.21

SVM Classifier in Case of Overlapping Distributions

P R E D I C T I V E A N A LY T I C S ◂

subject to y k(w T x k + b) ≥ 1 − ek , k = 1…n
ek ≥ 0.
The error variables ek are needed to allow for misclassifications.
The C hyperparameter in the objective function balances the importance of maximizing the margin versus minimizing the error on the
data. A high (low) value of C implies a higher (lower) risk of overfitting. We will come back to it in due course. Note that again a quadratic programming (QP) problem is obtained that can be solved using
Lagrangian optimization.
Finally, the nonlinear SVM classifier will first map the input data
to a higher dimensional feature space using some mapping ϕ( x ). This
is illustrated in Figure 3.22.
The SVM problem formulation now becomes:
Minimize

n
1 N 2
w
+
C
∑ i ∑ ei
2 i =1
i =1

subject to y k (w T ϕ( x k ) + b) ≥ 1 − ek , k = 1…n
ek ≥ 0.
When working out the Lagrangian optimization,24 it turns out that
the mapping ϕ( x ) is never explicitly needed, but only implicitly by
means of the kernel function K defined as follows: K( x k , x l ) = ϕ( x k )T ϕ( x l ).
Feature Space
X

Input Space
X

x →φ (x)

X X
X

X X

O
O

X X

O
X

X
O
O

X
X

K(x1,x2) = φ (x1)T φ (x2)

O
O

The Feature Space Mapping

O
O

Figure 3.22

X
X

O
O O O
X X

WT φ (xi) + b = 0

▸ A N A LY T I C S IN A BI G DATA WO RL D

Hence, the feature space does not need to be explicitly specified. The
nonlinear SVM classifier then becomes:
⎡ n
⎤
y( x ) = sign ⎢ ∑ α k yk K( x , x k ) + b ⎥
⎣ k =1
⎦
whereby α k are the Lagrangian multipliers stemming from the optimization. Support vectors will have nonzero α k since they are needed to
construct the classification line. All other observations have zero α k ,
which is often referred to as the sparseness property of SVMs. Different
types of kernel functions can be used. The most popular are:
■

Linear kernel: K( x , x k ) = x kT x

■

Polynomial kernel: K( x , x k ) = (1 + x kT x)d

■

Radial basis function (RBF) kernel: K( x , x k ) = exp{− || x − x k ||2 /σ 2 }

Empirical evidence has shown that the RBF kernel usually performs best, but note that it includes an extra parameter σ to be tuned.25
An SVM classifier can be very easily represented as a neural network, as depicted in Figure 3.23.
The hidden layer uses, for example, RBF activation functions,
whereas the output layer uses a linear activation function. Note that
the number of hidden neurons now corresponds to the number of
support vectors and follows automatically from the optimization. This
is in strong contrast to neural networks where the number of hidden
neurons needs to be tuned manually.

K(x,x1)

α1

K(x,x2)

α2

K(x,xns)

α ns

Figure 3.23

Representing an SVM Classifier as a Neural Network

P R E D I C T I V E A N A LY T I C S ◂

A key question to answer when building SVM classifiers is the tuning of the hyperparameters. For example, suppose one has an RBF
SVM that has two hyperparameters, C and σ. Both can be tuned using
the following procedure:26
■

Partition the data into 40/30/30 percent training, validation,
and test data.

■

Build an RBF SVM classifier for each (σ,C
C) combination from
the sets σ ∈ {0.5, 5, 10, 15, 25, 50, 100, 250, 500} and C ∈ {0.01,
0.05, 0.1, 0.5, 1, 5, 10, 50, 100, 500}.

■

Choose the (σ,C) combination with the best validation set performance.

■

Build an RBF SVM classifier with the optimal (σ,C) combination
on the combined training + validation data set.

■

Calculate the performance of the estimated RBF SVM classifier
on the test set.

In case of linear or polynomial kernels, a similar procedure can be
adopted.
SVMs can also be used for regression applications with a continuous target. The idea here is to find a function f(
f x) that has at most ε
deviation from the actual targets yi for all the training data, and is at
the same time as flat as possible. Hence, errors less (higher) than ε will
be tolerated (penalized). This is visualized in Figure 3.24.
n
N
Consider a training set: { x k , yk }k=1 with x k ∈ R and yk ∈ R

x
x
x

x
x

x
x
x

x
x

Loss
function

x
x

–ε
SVMs for Regression

+ε

▸ A N A LY T I C S IN A BI G DATA WO RL D

The SVM formulation then becomes:
Minimize

n
1 N 2
w i + C ∑(εk + εk* )
∑
2 i =1
i =1

subject to
y k − w T ϕ( x k ) − b ≤ ε + ε k
w T ϕ( x k ) + b − y k ≤ ε + ε k*
ε , ε k , ε k* ≥ 0.
The hyperparameter C determines the trade‐off between the flatness of f and the amount to which deviations larger than ε are tolerated. Note the feature space mapping ϕ( x ), which is also used here.
Using Lagrangian optimization, the resulting nonlinear regression
function becomes:
n

f ( x) = ∑(α k − α *k )K( x k , x ) + b ,
i =1

whereby α k and α *k represent the Lagrangian multipliers. The hyperparameters C and ε can be tuned using a procedure similar to the one
outlined for classification SVMs.
Just as with neural networks, SVMs have a universal approximation property. As an extra benefit, they do not require tuning of the
number of hidden neurons and are characterized by convex optimization. However, they are also very complex to be used in settings where
interpretability is important. Since an SVM can be represented as a
neural network (see Figure 3.23), one could use any of the rule extraction methods (decompositional, pedagogical) discussed in the section
on neural networks to make them more comprehensible.27 Also, two‐
stage models could be used to achieve this aim, whereby a second‐
stage SVM is estimated to correct for the errors of a simple (e.g., linear
or logistic regression) model.

ENSEMBLE METHODS
Ensemble methods aim at estimating multiple analytical models
instead of using only one. The idea here is that multiple models can
cover different parts of the data input space and, as such, complement
each other’s deficiencies. In order to successfully accomplish this, the

P R E D I C T I V E A N A LY T I C S ◂

analytical technique needs to be sensitive to changes in the underlying
data. This is especially the case for decision trees, and that’s why they
are commonly used in ensemble methods. In what follows, we will
discuss bagging, boosting, and random forests.

Bagging
Bagging (bootstrap aggregating) starts by taking B bootstraps from the
underlying sample.28 Note that a bootstrap is a sample with replacement
(see section on evaluating predictive models). The idea is then to build
a classifier (e.g., decision tree) for every bootstrap. For classification, a
new observation will be classified by letting all B classifiers vote, using, for
example, a majority voting scheme whereby ties are resolved arbitrarily.
For regression, the prediction is the average of the outcome of the B models (e.g., regression trees). Also note that here a standard error, and thus
confidence interval, can be calculated. The number of bootstraps B can
either be fixed (e.g., 30) or tuned via an independent validation data set.
The key element for bagging to be successful is the instability of the
analytical techniques. If perturbing the data set by means of the bootstrapping procedure can alter the model constructed, then bagging will
improve the accuracy.29

Boosting
Boosting works by estimating multiple models using a weighted sample
of the data.30 Starting from uniform weights, boosting will iteratively
reweight the data according to the classification error, whereby misclassified cases get higher weights. The idea here is that difficult observations should get more attention. Either the analytical technique can
directly work with weighted observations or, if not, we can just sample
a new data set according to the weight distribution. The final ensemble
model is then a weighted combination of all the individual models.
A popular implementation of this is the adaptive boosting/adaboost
procedure, which works as follows:
1. Given the following observations: (x1,y1), …, (xn, yn) where xi is
the attribute vector of observation i and yi ∈ {1,−1}
2. Initialize the weights as follows: W1(ii)=1/n, i = 1, …, n

▸ A N A LY T I C S IN A BI G DATA WO RL D

3. For t = 1…T
a. Train a weak classifier (e.g., decision tree) using the weights Wt
b. Get weak classifier Ct with classification error εt
1 ⎛ 1 − εt ⎞
c. Choose α t = ln ⎜
2 ⎝ ε t ⎟⎠
d. Update the weights as follows:
Wt (i) −αt
e if C t ( x) = y i
i. Wt +1(i) =
Zt
ii. Wt +1(i) =

Wt (i) αt
e if C t ( x) ≠ y i
Zt

⎛ T
⎞
4. Output the final ensemble model: E( x ) = sign ⎜ ∑(α t C t ( x ))⎟
⎝ t =1
⎠
Note that in the above procedure, T represents the number of boosting runs, αt measures the importance that is assigned to classifier Ct and
increases as εt gets smaller, Zt is a normalization factor needed to make
sure that the weights in step t make up a distribution and as such sum to
1, and C t ( x) represents the classification of the classifier built in step t for
observation x. Multiple loss functions may be used to calculate the error
εt, although the misclassification rate is undoubtedly the most popular.
In substep i of step d, it can be seen that correctly classified observations get lower weights, whereas substep ii assigns higher weights to
the incorrectly classified cases. Again, the number of boosting runs T
can be fixed or tuned using an independent validation set. Note that
various variants of this adaboost procedure exist, such as adaboost.M1,
adaboost.M2 (both for multiclass classification), and adaboost.R1 and
adaboost.R2 (both for regression).31 A key advantage of boosting is that
it is really easy to implement. A potential drawback is that there may be
a risk of overfitting to the hard (potentially noisy) examples in the data,
which will get higher weights as the algorithm proceeds.

Random Forests
Random forests was first introduced by Breiman.32 It creates a forest of
decision trees as follows:
1. Given a data set with n observations and N inputs
2. m = constant chosen on beforehand

P R E D I C T I V E A N A LY T I C S ◂

3. For t = 1,…, T
a. Take a bootstrap sample with n observations
b. Build a decision tree whereby for each node of the tree,
randomly choose m inputs on which to base the splitting
decision
c. Split on the best of this subset
d. Fully grow each tree without pruning
Common choices for m are 1, 2, or floor(log2(N) + 1), which is recommended. Random forests can be used with both classification trees
and regression trees. Key in this approach is the dissimilarity amongst
the base classifiers (i.e., decision trees), which is obtained by adopting
a bootstrapping procedure to select the training samples of the individual base classifiers, the selection of a random subset of attributes
at each node, and the strength of the individual base models. As such,
the diversity of the base classifiers creates an ensemble that is superior
in performance compared to the single models.
More recently, an alternative to random forests was proposed:
rotation forests. This ensemble technique takes the idea of random
forests one step further. It combines the idea of pooling a large number of decision trees built on a subset of the attributes and data, with
the application of principal component analysis prior to decision tree
building, explaining its name. Rotating the axes prior to model building was found to enhance base classifier accuracy at the expense of losing the ability of ranking individual attributes by their importance.33
Empirical evidence has shown that random forests can achieve excellent predictive performance at the cost of decreased comprehensibility.

MULTICLASS CLASSIFICATION TECHNIQUES
All of the techniques previously discussed can be easily extended to a
multiclass setting, whereby more than two target classes are available.

Multiclass Logistic Regression
When estimating a multiclass logistic regression model, one first needs
to know whether the target variable is nominal or ordinal. Examples

▸ A N A LY T I C S IN A BI G DATA WO RL D

of nominal targets could be predicting blood type and predicting voting
behavior. Examples of ordinal targets could be predicting credit ratings
and predicting income as high, medium, or low.
For nominal target variables, one of the target classes (say class K)
will be chosen as the base class as follows:
P(Y = 1| X 1 ,… , X N )
β1 +β1X +β1 X + β1N X N )
= e( 0 1 1 2 2
P(Y = K | X 1 ,… , X N )
P(Y = 2| X 1 ,…, X N )
β2 +β2 X +β2 X + β2N X N )
= e( 0 1 1 2 2
P(Y = K | X 1 ,…, X N )
...
P(Y = K − 1| X 1 ,… , X N ) (β0K −1 +β1K −1X1 +β2K −1X 2 + βKN−1X N )
=e
P(Y = K | X 1 ,… , X N )
Using the fact that all probabilities must sum to 1, one can obtain
the following:
β1 +β1X +β1 X + β1N X N )
e( 0 1 1 2 2
P(Y = 1| X 1 ,… , X N ) =
K −1 (βk +βk X +βk X + βkN X N )
1 + ∑ k =1 e 0 1 1 2 2
P(Y = 2| X 1 ,… , X N ) =
P(Y = K | X 1 ,… , X N ) =

β02 +β12 X1 +β22 X 2 + β2N X N

1 + ∑ k =1 e
K −1

1 + ∑ k =1

K −1

(

β0k +β1k X1 +βk2 X 2 +

)

βkN X N

)

1
β0k +β1k X1 +β2k X 2 + βkN X N )
(
e

The β parameters are then usually estimated using maximum
aposteriori estimation, which is an extension of maximum likelihood
estimation. As with binary logistic regression, the procedure comes
with standard errors, confidence intervals, and p‐values.
In case of ordinal targets, one could estimate a cumulative logistic
regression as follows:
P(Y ≤ 1) =
P(Y ≤ 2) =

1
1+ e

−θ1 +β1X1 + +β N X N

1+ e

−θ2 +β1X1 + +β N X N

P(Y ≤ K − 1) =

1
1 + e −θK −1+β1X1+ +βN X N

P R E D I C T I V E A N A LY T I C S ◂

or,
P(Y ≤ 1)
= e −θ1 +β1X1 + +βN X N
1 − P(Y ≤ 1)
P(Y ≤ 2)
= e −θ2 +β1X1+ +βN X N
1 − P(Y ≤ 2)
...
P(Y ≤ K − 1)
= e −θK −1+β1X1+ +βN X N
1 − P(Y ≤ K − 1)
Note that since P(Y ≤ K ) = 1 , θ K = +∞ .
The individual probabilities can then be obtained as follows:
P(Y = 1) = P(Y ≤ 1)
P(Y = 2) = P(Y ≤ 2) − P(Y ≤ 1)
...
P(Y = K ) = 1 − P(Y ≤ K − 1)
Also for this model, the β parameters can be estimated using a
maximum likelihood procedure.

Multiclass Decision Trees
Decision trees can be easily extended to a multiclass setting. For the
splitting decision, assuming K classes, the impurity criteria become:
K

Entropy(S) = − ∑ pk log2(pk )
k =1

Gini(S) = ∑ pk(1 − pk )
k =1

The stopping decision can be made in a similar way as for binary
target decision trees by using an independent validation data set. The
assignment decision then looks for the most prevalent class in each of
the leaf nodes.

Multiclass Neural Networks
A straightforward option for training a multiclass neural network
for K classes, is to create K output neurons, one for each class. An

▸ A N A LY T I C S IN A BI G DATA WO RL D

observation is then assigned to the output neuron with the highest
activation value (winner take all learning). Another option is to use a
softmax activation function.34

Multiclass Support Vector Machines
A common practice to estimate a multiclass support vector machine is
to map the multiclass classification problem to a set of binary classification problems. Two well‐known schemes here are one‐versus‐one and
one‐versus‐all coding.35
For K classes, one‐versus‐one coding estimates K(K − 1)/2 binary
SVM classifiers contrasting every possible pair of classes. Every classifier as such can cast a vote on the target class and the final classification is then the result of a (weighted) voting procedure. Ties are
resolved arbitrarily. This is illustrated in Figure 3.25.
For K classes, one‐versus‐all coding estimates K binary SVM classifiers each time contrasting one particular class against all the other
ones. A classification decision can then be made by assigning a particular observation to the class for which one of the binary classifiers
assigns the highest posterior probability. Ties are less likely to occur
with this scheme. This is illustrated in Figure 3.26.
Note that one‐versus‐one and one‐versus‐all are meta schemes
that can be used with other base classifiers as well.

Class is

x1
Figure 3.25

One‐versus‐One Coding for Multiclass Problems

P R E D I C T I V E A N A LY T I C S ◂

or other; p( ) = 0.92

or other; p( ) = 0.18

or other; p( ) = 0.30
Class is

x1
Figure 3.26

One‐versus‐All Coding for Multiclass Problems

EVALUATING PREDICTIVE MODELS
In this section, we will discuss how to evaluate the performance of
predictive models. First, we will discuss how to split up the data set.
This will be followed by a discussion of performance metrics.

Splitting Up the Data Set
When evaluating predictive models, two key decisions need to be
made. A first decision concerns the data set split up, which specifies
on what part of the data the performance will be measured. A second
decision concerns the performance metric. In what follows, we will
elaborate on both.
The decision how to split up the data set for performance measurement depends upon its size. In case of large data sets (say more
than 1,000 observations), the data can be split up into a training and
a test sample. The training sample (also called developmentt or estimation
sample) will be used to build the model, whereas the test sample (also
called the hold out sample) will be used to calculate its performance (see
Figure 3.27). There should be a strict separation between training and
test sample. Note that in case of decision trees or neural networks, the
validation sample should be part of the training sample because it is
actively being used during model development (i.e., to make the stopping decision).

▸ A N A LY T I C S IN A BI G DATA WO RL D

Train Data
Customer
John
Sarah
Sophie
David
Peter

Age Income Gender Good/Bad … Target
30 1,200
M
Bad
0
25
800
F
Good
1
52 2,200
F
Good
1
48 2,000
M
Bad
0
34 1,800
M
Good
1

Build Model

Data

Test Data
Customer
Emma
Will
Dan
Bob

Figure 3.27

Age
28
44
30
58

Income Gender … Good/Bad Score
1,000
F
Good
0.44
1,500
M
Bad
0.76
1,200
M
Good
0.18
2,400
M
Good
0.88

P(Good | age,income,...) =
1
1 + e −(0.10+0.50age+0.0034income +...)

Apply
Model

Training versus Test Sample Set Up for Performance Estimation

In the case of small data sets (say, less than 1,000 observations),
special schemes need to be adopted. A very popular scheme is cross‐
validation (see Figure 3.28). In cross‐validation, the data is split into
K folds (e.g., 10). A model is then trained on K − 1 training folds and
tested on the remaining validation fold. This is repeated for all possible
validation folds resulting in K performance estimates that can then be
averaged. Note also that a standard deviation and/or confidence interval can be calculated if desired. Common choices for K are 5 and 10. In
its most extreme case, cross‐validation becomes leave‐one‐out cross‐
validation whereby every observation is left out in turn and a model is
estimated on the remaining K − 1 observations. This gives K analytical
models in total. In stratified cross‐validation, special care is taken to
make sure the good/bad odds are the same in each fold.

.
.
.

Validation fold
Training fold
Figure 3.28

Cross‐Validation for Performance Measurement

P R E D I C T I V E A N A LY T I C S ◂

A key question to answer when doing cross‐validation is what
should be the final model that is being output from the procedure.
Because cross‐validation gives multiple models, this is not an obvious question. Of course, one could let all models collaborate in an
ensemble setup. A more pragmatic answer would be to, for example,
do leave‐one‐out cross‐validation and pick one of the models at random. Because the models differ up to one observation, they will be
quite similar anyway. Alternatively, one may also choose to build one
final model on all observations but report the performance coming out
of the cross‐validation procedure as the best independent estimate.
For small samples, one may also adopt bootstrapping procedures.
In bootstrapping, one takes samples with replacement from a data set D
(see Figure 3.29).
The probability that a customer is not sampled equals 1/n, with n
being the number of observations in the data set. Assuming a bootstrap
with n samples, the fraction of customers that is not sampled equals:
n

⎛1 − 1 ⎞ .
⎟
⎜⎝
n⎠
We then have:
n

1⎞
lim ⎛⎜ 1 − ⎟ = e −1 = 0.368
n→∞ ⎝
n⎠
whereby the approximation already works well for small values of n. So,
0.368 is the probability that a customer does not appear in the sample
and 0.632 is the probability that a customer does appear. If we then take
the bootstrap sample as the training set, and the test set as all samples in
D but not in the bootstrap, we can calculate the performance as follows:
Error estimate = 0.368 error(training) + 0.632 error(test),
whereby obviously a higher weight is being put on the test set performance.
D

C3
C1 C5
C4 C2

Figure 3.29

Bootstrapping

Bootstrap 1
Bootstrap 2

C3 C2 C5 C3 C2
C1 C4 C2 C1 C2

▸ A N A LY T I C S I N A BI G D ATA WO RL D

Table 3.3

Example Data Set for Performance Calculation

Churn

Score

John

Yes

0.72

Sophie

0.56

David

Yes

0.44

Emma

0.18

Bob

0.36

Performance Measures for Classification Models
Consider, for example, the following churn prediction example for
a five customer data set. The first column in Table 3.3 depicts the
true status, whereas the second column is the churn score as it
comes from a logistic regression, decision tree, neural network, and
so on.
One can now map the scores to a predicted classification label by
assuming a default cutoff of 0.5 as shown in Figure 3.30.
A confusion matrix can now be calculated and is shown in
Table 3.4.
Based upon this matrix, one can now calculate the following performance measures:
■

Classification accuracy = (TP + TN)/(TP + FP + FN + TN) = 3/5

■

Classification error = (FP + FN)/(TP + FP + FN + TN) = 2/5

■

Sensitivity = TP/(TP + FN) = 1/2

■

Specificity = TN/(FP + TN) = 2/3

However, note that all these classification measures depend on
the cut‐off. For example, for a cut off of 0 (1), classification accuracy

John
Sophie
David
Emma
Bob

Churn Score
Yes
0.72
No
0.56
Yes
0.44
No
0.18
No
0.36

Cutoff = 0.50

John
Sophie
David
Emma
Bob

Figure 3.30 Calculating Predictions Using a Cut‐Off

Churn
Yes
No
Yes
No
No

Score Predicted
0.72
Yes
0.56
Yes
0.44
No
0.18
No
0.36
No

P R E D I C T I V E A N A LY T I C S ◂

Table 3.4

The Confusion Matrix

Actual Status
Positive (churn)
Predicted status

Positive (churn)

True positive (John)

Negative (no churn)
False positive (Sophie)

Negative (no churn) False negative (David) True negative (Emma, Bob)

becomes 40 percent (60 percent), the error 60 percent (40 percent),
the sensitivity 100 percent (0), and the specificity 0 (100 percent).
It would be nice to have a performance measure that is independent from the cut‐off. One could construct a table with the sensitivity, specificity, and 1−specificity for various cut-offs as shown in
Table 3.5.
The receiver operating characteristic (ROC) curve then plots the
sensitivity versus 1−specificity as illustrated in Figure 3.31.36
Note that a perfect model has a sensitivity of 1 and a specificity of
1, and is thus represented by the upper left corner. The closer the curve
approaches this point, the better the performance. In Figure 3.31,
scorecard A has a better performance than scorecard B. A problem
arises, however, if the curves intersect. In this case, one can calculate the area under the ROC curve (AUC) as a performance metric.
The AUC provides a simple figure‐of‐merit for the performance of
the constructed classifier. The higher the AUC, the better the performance. The AUC is always bounded between 0 and 1 and can be
interpreted as a probability. In fact, it represents the probability that a
randomly chosen churner gets a higher score than a randomly chosen
nonchurner.37 Note that the diagonal represents a random scorecard
whereby sensitivity equals 1−specificity for all cut off points. Hence, a
Table 3.5 ROC Analysis

Cutoff
0

Sensitivity

Specificity

1−Specificity

0.01
0.02
….
0.99
1

▸ A N A LY T I C S IN A BI G DATA WO RL D

ROC Curve
Sensitivity

1
0.8
0.6
0.4
0.2
0

0.2

0.4

0.6

0.8

(1–Specificity)
Scorecard A
Figure 3.31

Random

Scorecard B

The Receiver Operating Characteristic Curve

good classifier should have an ROC above the diagonal and AUC bigger than 50%. Table 3.6 presents some AUC benchmarks for various
analytics applications.38
A lift curve is another important performance metric. It starts by
sorting the population from low score to high score. Suppose now
that in the top 10% lowest scores there are 60 percent bads, whereas
the total population has 10% bads. The lift value in the top decile
then becomes 60/10 percent or 6. In other words, the lift value represents the cumulative percentage of bads per decile, divided by the
overall population percentage of bads. Using no model, or a random
sorting, the bads would be equally spread across the entire range and
the lift value would always equal 1. Obviously, the lift curve always
decreases as one considers bigger deciles, until it will reach 1. This is
illustrated in Figure 3.32. Note that a lift curve can also be expressed
in a noncumulative way, and is also often summarized as the top
decile lift.
Table 3.6 Performance Benchmarks in Terms of AUC

Number of
Characteristics

AUC Ranges

Application credit scoring

10–15

70–85%

Behavioral credit scoring

10–15

80–90%

6–10

70–90%

10–15

70–90%

Application

Churn detection (telco)
Fraud detection (insurance)

P R E D I C T I V E A N A LY T I C S ◂

Scorecard
Baseline

5
4
3
2
1
0

Figure 3.32

50
60
70
% of Sorted Population

100

The Lift Curve

The cumulative accuracy profile (CAP), Lorenz, or power curve is
very closely related to the lift curve (see Figure 3.33). It also starts by
sorting the population from low score to high score and then measures
the cumulative percentage of bads for each decile on the Y‐axis. The
perfect model gives a linearly increasing curve up to the sample bad
rate and then flattens out. The diagonal again represents the random
model.
The CAP curve can be summarized in an Accuracy Ratio (AR) as
depicted in Figure 3.34.
The accuracy ratio is then defined as follows:
(Area below power curve for current model−Area below power
curve for random model)/
(Area below power curve for perfect model−Area below power
curve for random model)
A perfect model will thus have an AR of 1 and a random model an
AR of 0. Note that the accuracy ratio is also often referred to as the Gini
coefficient. There is also a linear relation between the AR and the AUC
as follows: AR = 2 * AUC−1.

Perfect Model

0.1

0.3

0.1

The Cumulative Accuracy Profile

Random model

0.2

0.4

0.6

0.8

1.2

Scorecard

Figure 3.33

Percentage of Bads

0.2

0.5

0.2

0.3

0.65

0.4

0.5

0.85

0.6

0.9

0.6

Percentage of sorted population

0.4

0.78

0.7

0.95

0.8

0.97

0.9

0.99

0.9

Perfect Model

Random model

Scorecard

P R E D I C T I V E A N A LY T I C S ◂

A
Current model

B
Perfect model
AR = B/(A + B)

Figure 3.34

Calculating the Accuracy Ratio

The Kolmogorov‐Smirnov distance is a separation measure calculating the maximum distance between the cumulative score distributions P(s | B) and P(s | G) defined as follows (see Figure 3.35):
P(s | G) = ∑ p( x | G)
x≤s

P(s | B) = ∑ p( x | B)
x ≤s

Note that by definition P(s | G) equals 1−sensitivity, and P(s | B)
equals the specificity. Hence, it can easily be verified that the KS distance can also be measured on an ROC graph. It fact, it is equal to the
maximum vertical distance between the ROC curve and the diagonal.

1
0.9
0.8

P(s|B)

0.7
0.6
0.5
0.4

KS distance

0.3
0.2
0.1

P(s|G)

0
Score
Figure 3.35

The Kolmogorov‐Smirnov Statistic

▸ A N A LY T I C S IN A BI G DATA WO RL D

Another performance measure is the Mahalanobis distance
between the score distributions, defined as follows:
M=

| μG − μ B |
,
σ

whereby μG (μ B ) represents the mean score of the goods (bads) and
σ the pooled standard deviation. Obviously, a high Mahalanobis
distance is preferred because it means both score distributions are
well separated. Closely related is the divergence metric, calculated
as follows:
D=

(μ G − μ B )2
1 2
(σ + σ 2B )
2 G

be adopted. Figure 3.36 presents an example of a multiclass confusion
matrix.
The on‐diagonal elements represented in gray correspond to the
correct classifications. Off‐diagonal elements represent errors. Note,
however, that not all errors have equal impact. Given the ordinal
nature of the target variable, the further away from the diagonal,
the bigger the impact of the error. For example, in Figure 3.36,
there are 34 C+ observations predicted as C, which is not as bad as
the one C+ predicted as D. One could summarize this in a notch difference graph that depicts the cumulative accuracy for increasing
notch differences. Figure 3.37 gives an example of a notch difference graph.
At the 0 notch difference level, the performance equals about
35 percent, which may not seem very good. However, by allowing
for a one‐notch difference, the performance almost doubles to around
75 percent, which is a lot better!
The AUC can be generalized to the multiclass setting by plotting an ROC graph for each class against all other classes, calculating
the AUC, and taking the overall average. Another option is to calculate an AUC for each possible class comparison and then take the
average.39

A+
A
A–
B+
B
B–
C+
C
C–
D+
D
D–
E+
E
E–

Figure 3.36

True

A
0
14
0
7
2
0
0
1
0
0
0
0
0
0
0

A–
0
0
0
0
2
0
0
0
0
0
0
0
0
0
0

B+
0
4
2
50
54
6
13
5
2
1
1
0
0
0
0

B
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

Confusion Matrix for a Multiclass Example

A+
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

B–
0
0
0
0
12
6
3
1
0
0
0
0
0
0
0

C
0
0
0
5
17
6
34
151
8
20
9
0
4
4
0

Predicted
C+
0
0
0
5
58
10
99
37
3
7
2
0
0
0
0

C–
0
0
0
0
0
0
0
1
3
2
1
0
0
0
0

D+
0
0
0
1
5
1
11
12
4
131
32
0
12
4
0

D
0
0
0
0
1
0
1
4
0
24
122
0
26
9
0

D–
0
0
0
0
0
0
0
0
0
0
0
2
0
0
0

E+
0
0
0
0
1
0
0
0
0
3
13
1
37
5
0

E
0
0
0
0
0
0
0
0
0
1
7
0
5
48
0

E–
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

100

Figure 3.37

Cumulative Accuracy (%)

A Cumulative Notch Difference Graph

3
Notch Difference

P R E D I C T I V E A N A LY T I C S ◂

Performance Measures for Regression Models
Multiple measures exist to calculate the performance of regression
models. A first key metric is the R‐squared, defined as follows:

∑
= 1−
∑

( yi − yˆ i )2

i =1
n

( yi − y )2

i =1

whereby y i is the true value, yˆ i the predicted value, and y the average.
The R 2 always varies between 0 and 1, and higher values are to be preferred. Two other popular measures are the mean squared error (MSE)
and mean absolute deviation (MAD), defined as follows:

∑
MSE =

∑
MAD =

( y i − yˆ i )2

i =1

n
i =1

| yi − yˆ i |
n

a visual representation of model performance (see Figure 3.38). The
more the plot approaches a straight line through the origin, the better
the performance of the model. It can be summarized by calculating the
Pearson correlation coefficient.
50
P
r
e
d
i
c
t
e
d

45
40
35
30
25
20

15
C
L 10
V
5
0

CLV
Figure 3.38

Scatter Plot for Measuring Model Performance

▸ A N A LY T I C S I N A BI G D ATA WO RL D

NOTES
1. T. Hastie, R. Tibshirani, and J. Friedman, Elements of Statistical Learning: Data Mining,
Inference and Prediction (Springer‐Verlag, Berlin, Germany, 2001).
2. S. Viaene et al., “A Comparison of State‐of‐the‐Art Classification Techniques for
Expert Automobile Insurance Fraud Detection.” Special issue, Journal of Risk and
Insurance 69, no. 3 (2002): 433–443.
3. S. Gupta et al., “Modeling Customer Lifetime Value,” Journal of Service Research 9,
no. 2 (2006): 139–155; N. Glady, C. Croux, and B. Baesens, “Modeling Churn Using
Customer Lifetime Value,” European Journal of Operational Research 197, no. 1 (2009):
402–411.
4. T. Van Gestel and B. Baesens, Credit Risk Management: Basic Concepts: Financial Risk
Components, Rating Analysis, Models, Economic and Regulatory Capitall (Oxford University
Press, 2009); G. Loterman et al., “Benchmarking Regression Algorithms for Loss
Given Default Modeling,” International Journal of Forecasting 28, no. 1 (2012): 161–
170; E. Tobback et al., “Forecasting Loss Given Default Models: Impact of Account
Characteristics and the Macroeconomic State,” Journal of the Operational Research
Society, forthcoming 2014.
5. D. W. Dwyer, A. Kocagil, and R. Stein, The Moody’s KMV EDF™ RiskCalc™ v3.1 Model
Next Generation Technology for Predicting Private Firm Credit Risk
k (White paper, 2004).
6. R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification (John Wiley & Sons,
Hoboken, New Jersey, US, 2001).
7. J. R. Quinlan, C4.5 Programs for Machine Learning (Morgan Kauffman Publishers,
Burlington, Massachusetts, United States, 1993).
8. L. Breiman et al., Classification and Regression Trees (Monterey, CA: Wadsworth &
Brooks/Cole Advanced Books & Software, 1984).
9. J. A. Hartigan, Clustering Algorithms (New York: John Wiley & Sons, 1975).
10. C. M. Bishop, Neural Networks for Pattern Recognition (Oxford University Press, Oxford,
England, 1995); J. M. Zurada, Introduction to Artificial Neural Systems (Boston: PWS
Publishing, 1992).
11. B. Baesens et al., “Bayesian Neural Network Learning for Repeat Purchase Modelling in Direct Marketing,” European Journal of Operational Research 138, no. 1 (2002):
191–211.
12. K. Hornik, M. Stinchcombe, and H. White, “Multilayer Feedforward Networks Are
Universal Approximators,” Neural Networks 2, no. 5 (1989): 359–366.
13. See C. M. Bishop, Neural Networks for Pattern Recognition (Oxford University Press,
Oxford, England, 1995) for more details.
14. J. Moody and J. Utans. “Architecture Selection Strategies for Neural Networks:
Application to Corporate Bond Rating Prediction,” in Neural Networks in the Capital
Markets, A. N. Refenes (editor) (New York: John Wiley & Sons, 1994).
15. P. L. Bartlett, “For Valid Generalization, the Size of the Weights Is More Important
than the Size of the Network,” in Advances in Neural Information Processing Systems 9,
ed. M. C, Mozer, M. I. Jordan, and T. Petsche (Cambridge, MA: MIT Press, 1997),
134–140.
16. B. Baesens, D. et al., “White Box Nonlinear Prediction Models.” Special issue, IEEE
Transactions on Neural Networks 22, no. 12 (2011): 2406–2408.
17. B. Baesens, “Developing Intelligent Systems for Credit Scoring using Machine Learning Techniques” (PhD thesis, Katholieke Universiteit Leuven, 2003); B. Baesens et al.,

P R E D I C T I V E A N A LY T I C S ◂

18.

19.

20.

21.

22.
23.

24.

25.
26.
27.

28.
29.
30.

31.

“Using Neural Network Rule Extraction and Decision Tables for Credit‐Risk Evaluation,” Management Science 49, no. 3 (2003): 312–329; R. Setiono, B. Baesens, and C.
Mues, “A Note on Knowledge Discovery Using Neural Networks and Its Application
to Credit Card Screening,” European Journal of Operational Research 192, no. 1 (2009):
326–332.
H. Lu, R. Setiono, and H. Liu, “NeuroRule: A Connectionist Approach to Data
Mining,” in Proceedings of 21st International Conference on Very Large Data Bases (Zurich,
Switzerland, Morgan Kaufmann, 1995), 478–489.
M. Craven and J. Shavlik, “Extracting Tree‐Structured Representations of Trained
Networks,” in Advances in Neural Information Processing Systems (Cambridge, MA: MIT
Press, 1996).
T. Van Gestel et al., “Linear and Nonlinear Credit Scoring by Combining Logistic
Regression and Support Vector Machines,” Journal of Credit Riskk 1, no. 4 (2005); T.
Van Gestel et al., “A Process Model to Develop an Internal Rating System: Sovereign
Credit Ratings,” Decision Support Systems 42, no. 2 (2006): 1131–1151.
N. Cristianini and J. S. Taylor, An Introduction to Support Vector Machines and Other
Kernel‐based Learning Methods (Cambridge University Press, 2000); B. Schölkopf and
A. Smola, Learning with Kernels (Cambridge, MA: MIT Press, 2001); V. Vapnik, The
Nature of Statistical Learning Theory (New York: Springer‐Verlag, 1995).
O. L. Mangasarian, “Linear and Non‐linear Separation of Patterns by Linear Programming,” Operations Research 13, May‐June (1965): 444–452.
N. Cristianini and J. S. Taylor, An Introduction to Support Vector Machines and Other
Kernel‐based Learning Methods (Cambridge University Press, 2000); B. Schölkopf and
A. Smola, Learning with Kernels (Cambridge, MA: MIT Press, 2001); V. Vapnik, The
Nature of Statistical Learning Theory (New York: Springer‐Verlag, 1995).
N. Cristianini and J. S. Taylor, An Introduction to Support Vector Machines and Other
Kernel‐based Learning Methods (Cambridge University Press, 2000); B. Schölkopf and
A. Smola, Learning with Kernels (Cambridge, MA: MIT Press, 2001); V. Vapnik, The
Nature of Statistical Learning Theory (New York: Springer‐Verlag, 1995).
T. Van Gestel et al., “Benchmarking Least Squares Support Vector Machine Classifiers,” Machine Learning 54, no. 1 (2004): 5–32.
Ibid.
D. Martens et al., “Comprehensible Credit Scoring Models Using Rule Extraction
From Support Vector Machines,” European Journal of Operational Research 183 (2007):
1466–1476; D. Martens, B. Baesens, and T. Van Gestel, “Decompositional Rule
Extraction from Support Vector Machines by Active Learning,” IEEE Transactions on
Knowledge and Data Engineering 21, no. 1, (2009): 178–191.
L. Breiman, “Bagging Predictors,” Machine Learning 24, no. 2 (1996): 123–140.
Ibid.
Y. Freund and R. E. Schapire, “A Decision‐Theoretic Generalization of On‐Line
Learning and an Application to Boosting,” Journal of Computer and System Sciences
55, no. 1 (1997): 119–139; Y. Freund and R. E. Schapire, “A Short Introduction
to Boosting,” Journal of Japanese Society for Artificial Intelligence 14, no. 5 (1999):
771–780.
See Y. Freund and R. E. Schapire, “A Decision‐Theoretic Generalization of On‐Line
Learning and an Application to Boosting,” Journal of Computer and System Sciences
55, no. 1 (1997): 119–139, and Y. Freund and R. E. Schapire, “A Short Introduction to Boosting,” Journal of Japanese Society for Artificial Intelligence 14, no. 5 (1999):
771–780, for more details.

▸ A N A LY T I C S IN A BI G DATA WO RL D

32. L. Breiman, “Random Forests,” Machine Learning 45, no. 1 (2001): 5–32.
33. J. J. Rodriguez, L. I. Kuncheva, and C. J. Alonso, “Rotation Forest: A New Classifier
Ensemble Method,” IEEE Transactions on Pattern Analysis and Machine Intelligence 28,
no. 10 (2006): 1619–1630.
34. C. M. Bishop, Neural Networks for Pattern Recognition (Oxford University Press, Oxford,
England, 1995).
35. T. Van Gestel, “From Linear to Kernel Based Methods for Classification, Modelling
and Prediction” (PhD Thesis, Katholieke Universiteit Leuven, 2002).
36. T. Fawcett, “ROC Graphs: Notes and Practical Considerations for Researchers,” HP
Labs Tech Report HPL‐2003–4, HP Laboratories, Palo Alto, US (2003).
37. E. R. Delong, D. M. Delong, and D. L. Clarke‐Pearson, “Comparing the Areas Under
Two or More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach,” Biometrics 44 (1988): 837–845; J. A, Hanley and B. J. McNeil, “The
Meaning and Use of Area under the ROC Curve,” Radiology 143 (1982): 29–36.
38. B. Baesens et al., “Benchmarking State of the Art Classification Algorithms for
Credit Scoring,” Journal of the Operational Research Society 54, no. 6 (2003): 627–635;
W. Verbeke et al., “New Insights into Churn Prediction in the Telecommunication
Sector: A Profit Driven Data Mining Approach,” European Journal of Operational
Research 218, no. 1 (2012): 211–229.
39. D. Hand and R. J. Till, “A Simple Generalization of the Area under the ROC Curve to
Multiple Class Classification Problems,” Machine Learning 45, no. 2 (2001): 171–186.

C H A P T E R

Descriptive
Analytics

n descriptive analytics, the aim is to describe patterns of customer
behavior. Contrary to predictive analytics, there is no real target
variable (e.g., churn or fraud indicator) available. Hence, descriptive
analytics is often referred to as unsupervised learning because there is no
target variable to steer the learning process. The three most common
types of descriptive analytics are summarized in Table 4.1.

ASSOCIATION RULES
In this section, we will address how to mine association rules from
data. First, the basic setting will be discussed. This will be followed by a
discussion of support and confidence, which are two key measures for
association rule mining. Next, we will zoom into the association rule
mining process. The lift measure will then be introduced. The section
will be concluded by discussing post processing, extensions, and various applications of association rules.

Basic Setting
Association rules typically start from a database of transactions, D. Each
transaction consists of a transaction identifier and a set of items (e.g.,

▸ A N A LY T I C S IN A BI G DATA WO RL D

Table 4.1

Examples of Descriptive Analytics

Type of Descriptive
Analytics
Explanation
Association rules

Example

Detect frequently
occurring
patterns between
items

Detecting what products are frequently purchased
together in a supermarket context
Detecting what words frequently co‐occur in a
text document
Detecting what elective courses are frequently
chosen together in a university setting

Sequence rules

Detect
sequences of
events

Detecting sequences of purchase behavior in a
supermarket context
Detecting sequences of web page visits in a web
mining context
Detecting sequences of words in a text document

Clustering

Detect
homogeneous
segments of
observations

Differentiate between brands in a marketing
portfolio
Segment customer population for targeted
marketing

products, Web pages, courses) {i1, i2, …, in} selected from all possible
items (I).
I Table 4.2 gives an example of a transactions database in a
supermarket setting.
An association rule is then an implication of the form X ⇒ Y,
whereby X ⊂ I, Y ⊂ I and X ∩ Y = ∅. X is referred to as the rule

Table 4.2

Example Transaction Data Set

Transaction Identifier

Items

Beer, milk, diapers, baby food

Coke, beer, diapers

Cigarettes, diapers, baby food

Chocolates, diapers, milk, apples

Tomatoes, water, apples, beer

Spaghetti, diapers, baby food, beer

Water, beer, baby food

Diapers, baby food, spaghetti

Baby food, beer, diapers, milk

Apples, wine, baby food

D E S C R I P T I V E A N A LY T I C S ◂

antecedent, whereas Y is referred to as the rule consequent. Examples
of association rules are:
■

If a customer has a car loan and car insurance, then the customer has a checking account in 80% of the cases.

■

If a customer buys spaghetti, then the customer buys red wine
in 70 percent of the cases.

■

If a customer visits web page A, then the customer will visit web
page B in 90% of the cases.

It is hereby important to note that association rules are stochastic
in nature, which means they should not be interpreted as a universal truth and are characterized by statistical measures quantifying the
strength of the association. Also, the rules measure correlational associations and should not be interpreted in a causal way.

Support and Confidence
Support and confidence are two key measures to quantify the strength
of an association rule. The support of an item set is defined as the percentage of total transactions in the database that contains the item set.
Hence, the rule X ⇒ Y has support (s) if 100s% of the transactions in D
contain X ∪ Y. It can be formally defined as follows:
support(X ∪ Y ) =

number of transactions supporting (X ∪ Y )
total number of transactions

When considering the transaction database in Table 4.2, the association
rule baby food and diapers ⇒ beer has support 3/10 or 30 percent.
A frequent item set is one for which the support is higher than a
threshold (minsup) that is typically specified upfront by the business
user or data analyst. A lower (higher) support will obviously generate
more (less) frequent item sets. The confidence measures the strength of
the association and is defined as the conditional probability of the rule
consequent, given the rule antecedent. The rule X ⇒ Y has confidence
(c) if 100c% of the transactions in D that contain X also contain Y.
It can be formally defined as follows:
confidence(X → Y ) = P(Y | X ) =

support(X ∪ Y )
support(X )

▸ A N A LY T I C S IN A BI G DATA WO RL D

Again, the data analyst has to specify a minimum confidence (minconf) in order for an association rule to be considered interesting.
When considering Table 4.2, the association rule baby food and
diapers ⇒ beer has confidence 3/5 or 60 percent.

Association Rule Mining
Mining association rules from data is essentially a two‐step process as
follows:
1. Identification of all item sets having support above minsup (i.e.,
“frequent” item sets)
2. Discovery of all derived association rules having confidence
above minconf
As said before, both minsup and minconf need to be specified
beforehand by the data analyst. The first step is typically performed
using the Apriori algorithm.1 The basic notion of a priorii states that every
subset of a frequent item set is frequent as well or, conversely, every
superset of an infrequent item set is infrequent. This implies that candidate item sets with k items can be found by pairwise joining frequent
item sets with k − 1 items and deleting those sets that have infrequent
subsets. Thanks to this property, the number of candidate subsets to
be evaluated can be decreased, which will substantially improve the
performance of the algorithm because fewer databases passes will be
required. The Apriori algorithm is illustrated in Figure 4.1.
Once the frequent item sets have been found, the association rules
can be generated in a straightforward way, as follows:
■

For each frequent item set k, generate all nonempty subsets of k

■

For every nonempty subset s of k, output the rule s ⇒ k − s if the
confidence > minconf

Note that the confidence can be easily computed using the support
values that were obtained during the frequent item set mining.
For the frequent item set {baby food, diapers, beer}, the following
association rules can be derived:
diapers, beer ⇒ baby food [conff = 75%]
baby food, beer ⇒ diapers [conff = 75%]

D E S C R I P T I V E A N A LY T I C S ◂

Database

L1
TID

Items

Itemsets

100
200
300
400

1, 3, 4
2, 3, 5
1, 2, 3, 5
2, 5

{1}
{2}
{3}
{5}

Support
2/4
3/4
3/4
3/4

Minsup = 50%

Itemsets
{1, 2}
{1, 3}
{1, 5}
{2, 3}
{2, 5}
{3, 5}

Support

Itemsets

1/4
2/4
1/4
2/4
3/4
2/4

{1, 3}
{2, 3}
{2, 5}
{3, 5}

Support
2/4
2/4
3/4
2/4

C3
Itemsets

Support

Itemsets

Support

{2, 3, 5}

2/4

{2, 3, 5}

2/4

{1,3} and {2,3} give
{1,2,3}, but because {1,2}
is not frequent, you do not
have to consider it!

Result = {{1},{2},{3},{5},{1,3},{2,3},{2,5},{3,5},{2,3,5}}
Figure 4.1 The Apriori Algorithm

baby food, diapers ⇒ beer [conff = 60%]
beer ⇒ baby food and diapers [conff = 50%]
baby food ⇒ diapers and beer [conff = 43%]
diapers ⇒ baby food and beer [conff = 43%]
If the minconf is set to 70 percent, only the first two association
rules will be kept for further analysis.

The Lift Measure
Table 4.3 provides an example from a supermarket transactions database to illustrate the lift measure.
Let’s now consider the association rule tea ⇒ coffee. The support
of this rule is 100/1,000, or 10 percent. The confidence of the rule is
Table 4.3

Coffee
Not coffee
Total

The Lift Measure

Tea

Not Tea

150

750

Total
900

100

200

800

1,000

▸ A N A LY T I C S IN A BI G DATA WO RL D

150/200, or 75 percent. At first sight, this association rule seems very
appealing given its high confidence. However, closer inspection reveals
that the prior probability of buying coffee equals 900/1,000, or 90 percent. Hence, a customer who buys tea is less likely to buy coffee than a
customer about whom we have no information. The lift, also referred
to as the interestingness measure, takes this into account by incorporating
the prior probability of the rule consequent, as follows:
Lift(X → Y ) =

support(X ∪ Y )
support(X ) i support(Y )

A lift value less (larger) than 1 indicates a negative (positive)
dependence or substitution (complementary) effect. In our example,
the lift value equals 0.89, which clearly indicates the expected substitution effect between coffee and tea.

Post Processing Association Rules
Typically, an association rule mining exercise will yield lots of association rules such that post processing will become a key activity. Example steps that can be considered here are:
■

Filter out the trivial rules that contain already known patterns
(e.g., buying spaghetti and spaghetti sauce). This should be done
in collaboration with a business expert.

■

Perform a sensitivity analysis by varying the minsup and minconf values. Especially for rare but profitable items (e.g., Rolex
watches), it could be interesting to lower the minsup value and
find the interesting associations.

■

Use appropriate visualization facilities (e.g., OLAP based) to find
the unexpected rules that might represent novel and actionable
behavior in the data.

■

Measure the economic impact (e.g., profit, cost) of the association rules.

Association Rule Extensions
Since the introduction of association rules, various extensions have
been proposed. A first extension would be to include item quantities

D E S C R I P T I V E A N A LY T I C S ◂

...
Beverages
Non-Gassy
Drinks

Milk

UHT Milk

Strawberry
Milk

Chocolate
Milk

Carbonated
Drinks

Beer

Fresh Milk

Vanilla
Milk

Plain
Milk

Figure 4.2 A Product Taxonomy for Association Rule Mining

and/or price. This can be easily accomplished by adding discretized
quantitative variables (e.g., three bottles of milk) to the transaction
data set and mine the frequent item sets using the Apriori algorithm.
Another extension is to also include the absence of items. Also, this
can be achieved by adding the absence of items to the transactions
data set and again mine using the Apriori algorithm. Finally, multilevel
association rules mine association rules at different concept levels of a
product taxonomy, as illustrated in Figure 4.2.2 A similar approach can
again be followed here by adding taxonomy information to the transactions data set. Note that different support levels may have to be set
for different levels of the product taxonomy.

Applications of Association Rules
The most popular application of association rules is market basket
analysis. The aim here is to detect which products or services are
frequently purchased together by analyzing market baskets. Finding
these associations can have important implications for targeted marketing (e.g., next best offer), product bundling, store and shelf layout,
and/or catalog design.
Another popular application is recommender systems. These are
the systems adopted by companies such as Amazon and Netflix to give
a recommendation based on past purchases and/or browsing behavior.

▸ A N A LY T I C S IN A BI G DATA WO RL D

SEQUENCE RULES
Given a database D of customer transactions, the problem of mining
sequential rules is to find the maximal sequences among all sequences
that have certain user‐specified minimum support and confidence. An
example could be a sequence of web page visits in a web analytics
setting, as follows:
Home page ⇒ Electronics ⇒ Cameras and Camcorders ⇒ Digital
Cameras ⇒ Shopping cart ⇒ Order confirmation ⇒ Return to shopping
It is important to note that a transaction time or sequence field
will now be included in the analysis. Whereas association rules are
concerned about what items appear together at the same time (intratransaction patterns), sequence rules are concerned about what
items appear at different times (intertransaction patterns). To mine
the sequence rules, one can again make use of the a priorii property
because if a sequential pattern of length k is infrequent, its supersets of
length k + 1 cannot be frequent.
Consider the following example of a transactions data set in a web
analytics setting (see Table 4.4). The letters A, B, C, … refer to web pages.

Table 4.4

Example Transactions Data Set for Sequence Rule Mining

Session ID
1
1
1
2
2
3
3
3
4
4
4
5
5
5

Page

Sequence

A
B
C
B
C
A
C
D
A
B
D
D
C
A

1
2
3
1
2
1
2
3
1
2
3
1
1
1

D E S C R I P T I V E A N A LY T I C S ◂

A sequential version can then be obtained as follows:
Session 1: A, B, C
Session 2: B, C
Session 3: A, C, D
Session 4: A, B, D
Session 5: D, C, A
One can now calculate the support in two different ways. Consider, for example, the sequence rule A ⇒ C. A first approach would
be to calculate the support whereby the consequent can appear in any
subsequent stage of the sequence. In this case, the support becomes
2/5 (40%). Another approach would be to only consider sessions in
which the consequent appears right after the antecedent. In this case,
the support becomes 1/5 (20%). A similar reasoning can now be followed for the confidence, which can then be 2/4 (50%) or 1/4 (25%),
respectively.
Remember that the confidence of a rule A1 ⇒ A2 is defined as the
probability P(A2 | A1) = support(A1 ∪ A2)/support(A1). For a rule with
multiple items, A1 ⇒ A2 ⇒ … An–1 ⇒ An, the confidence is defined as
P(An | A1, A2, …, An–1), or support(A1 ∪ A2 ∪ … ∪ An–1 ∪ An)/support
(A1 ∪ A2 ∪ … ∪ An–1).

SEGMENTATION
The aim of segmentation is to split up a set of customer observations into segments such that the homogeneity within a segment is
maximized (cohesive) and the heterogeneity between segments is
maximized (separated). Popular applications include:
■

Understanding a customer population (e.g., targeted marketing
or advertising [mass customization])

■

Efficiently allocating marketing resources

■

Differentiating between brands in a portfolio

■

Identifying the most profitable customers

■

Identifying shopping patterns

■

Identifying the need for new products

▸ A N A LY T I C S IN A BI G DATA WO RL D

Clustering

Hierarchical

Agglomerative

Nonhierarchical

Divisive

SOM

k‐means

Figure 4.3 Hierarchical versus Nonhierarchical Clustering Techniques

Various types of clustering data can be used, such as demographic,
lifestyle, attitudinal, behavioral, RFM, acquisitional, social network,
and so on.
Clustering techniques can be categorized as either hierarchical or
nonhierarchical (see Figure 4.3).

Hierarchical Clustering
In what follows, we will first discuss hierarchical clustering. Divisive
hierarchical clustering starts from the whole data set in one cluster, and then breaks this up in each time smaller clusters until one
observation per cluster remains (right to left in Figure 4.4). Agglomerative clustering works the other way around, starting from all
Divisive
Step 4
C1
C2

Step 3

Step 2

Step 0
C1
C2
C3
C4
C5

C1
C2

C3
C4

C4
C5

Step 0

Step 1

Step 2

C3
C4
C5

Step 3

Step 4

Agglomerative
Divisive versus Agglomerative Hierarchical Clustering

D E S C R I P T I V E A N A LY T I C S ◂

Recency
20

Manhattan

30
Figure 4.5

Manhattan

Monetary

Euclidean versus Manhattan Distance

observations in one cluster and continuing to merge the ones that
are most similar until all observations make up one big cluster (left to
right in Figure 4.4).
In order to decide on the merger or splitting, a similarity rule is
needed. Examples of popular similarity rules are the Euclidean distance
and Manhattan (city block) distance. For the example in Figure 4.5,
both are calculated as follows:
Euclidean : (50 − 30)2 + (20 − 10)2 = 22
Manhattan: ⎢50 − 30 ⎢+ ⎢20 − 10 ⎢= 30
It is obvious that the Euclidean distance will always be shorter
than the Manhattan distance.
Various schemes can now be adopted to calculate the distance
between two clusters (see Figure 4.6). The single linkage method
Single linkage

Complete linkage
Average linkage

Centroid method

Figure 4.6

Calculating Distances between Clusters

▸ A N A LY T I C S IN A BI G DATA WO RL D

Chicken

Parrot
Pigeon

Duck

Canary

Owl

Eagle
6

Figure 4.7 Example for Clustering Birds
The numbers indicate the clustering steps.

defines the distance between two clusters as the shortest possible
distance, or the distance between the two most similar objects. The
complete linkage method defines the distance between two clusters as
the biggest distance, or the distance between the two most dissimilar
objects. The average linkage method calculates the average of all possible distances. The centroid method calculates the distance between
the centroids of both clusters. Finally, Ward’s method merges the pair
of clusters that leads to the minimum increase in total within‐cluster
variance after merging.
In order to decide on the optimal number of clusters, one could
use a dendrogram or scree plot. A dendrogram is a tree‐like diagram
that records the sequences of merges. The vertical (or horizontal
scale) then gives the distance between two clusters amalgamated. One
can then cut the dendrogram at the desired level to find the optimal
clustering. This is illustrated in Figure 4.7 and Figure 4.8 for a birds
clustering example. A scree plot is a plot of the distance at which clusters are merged. The elbow point then indicates the optimal clustering.
This is illustrated in Figure 4.9.

D E S C R I P T I V E A N A LY T I C S ◂

4
3
2

Chicken Duck Pigeon

Parrot

Canary

Owl

Eagle

Figure 4.8 Dendrogram for Birds Example
The black line indicates the optimal clustering.

K‐Means Clustering
K‐means clustering is a nonhierarchical procedure that works along
the following steps:
1. Select k observations as initial cluster centroids (seeds).
2. Assign each observation to the cluster that has the closest centroid (for example, in Euclidean sense).
3. When all observations have been assigned, recalculate the positions of the k centroids.
4. Repeat until the cluster centroids no longer change.

Distance

A key requirement here is that the number of clusters, k, needs to
be specified before the start of the analysis. It is also advised to try out
different seeds to verify the stability of the clustering solution.

Number of Clusters
Figure 4.9 Scree Plot for Clustering

100

▸ A N A LY T I C S I N A BI G DATA WO RL D

Self‐Organizing Maps
A self‐organizing map (SOM) is an unsupervised learning algorithm
that allows you to visualize and cluster high‐dimensional data on a
low‐dimensional grid of neurons.3 An SOM is a feedforward neural
network with two layers. The neurons from the output layer are usually ordered in a two‐dimensional rectangular or hexagonal grid (see
Figure 4.10). For the former, every neuron has at most eight neighbors, whereas for the latter every neuron has at most six neighbors.
Each input is connected to all neurons in the output layer with
weights w = [w1, …, wN], with N the number of variables. All weights
are randomly initialized. When a training vector x is presented, the
weight vector wc of each neuron c is compared with x, using, for
example, the Euclidean distance metric (beware to standardize the
data first):
d( x , w c ) =

∑( x

− w ci )2

i =1

x in Euclidean sense is called
the best matching unit (BMU). The weight vector of the BMU
and its neighbors in the grid are then adapted using the following
learning rule:
w i (t + 1) = w i (t + 1) + hci (t)[ x(t) − w i (t)]
whereby t represents the time index during training and hcii(t) defines
the neighborhood of the BMU c, specifying the region of influence. The

Rectangular SOM Grid

Figure 4.10

Hexagonal SOM Grid

Rectangular versus Hexagonal SOM Grid

D E S C R I P T I V E A N A LY T I C S ◂

101

neighborhood function hcii(t) should be a nonincreasing function of
time and the distance from the BMU. Some popular choices are:
⎛ r −r 2⎞
hci (t) = α(t)exp ⎜ − c 2 i ⎟
⎝ 2σ (t) ⎠
hci (t ) = α(t ) if rc − ri

≤ threshold,0 otherwise,

whereby rc and ri represent the location of the BMU and neuron i on
the map, σ2(t) represents the decreasing radius, and 0 ≤ α(t) ≤ 1, the
learning rate (e.g., α(tt) = A/(t + B), α(t) = exp(––Att)). The decreasing
learning rate and radius will give a stable map after a certain amount
of training. Training is stopped when the BMUs remain stable, or after
a fixed number of iterations (e.g., 500 times the number of SOM neurons). The neurons will then move more and more toward the input
observations and interesting segments will emerge.
SOMs can be visualized by means of a U‐matrix or component
plane.
■

A U (unified distance)‐matrix essentially superimposes a height
Z dimension on top of each neuron visualizing the average distance between the neuron and its neighbors, whereby typically
dark colors indicate a large distance and can be interpreted as
cluster boundaries.

■

A component plane visualizes the weights between each specific input variable and its output neurons, and as such provides
a visual overview of the relative contribution of each input attribute to the output neurons.

Figure 4.11 provides an SOM example for clustering countries
based on a corruption perception index (CPI). This is a score between 0
(highly corrupt) and 10 (highly clean) assigned to each country in the
world. The CPI is combined with demographic and macroeconomic
information for the years 1996, 2000, and 2004. Upper case countries
(e.g., BEL) denote the situation in 2004, lowercase (e.g., bel) in 2000,
and sentence case (e.g., Bel) in 1996. It can be seen that many of the
European countries are situated in the upper right corner of the map.

▸ A N A LY T I C S I N A BI G D ATA WO RL D

102

sgp
sgp
SGP

ISR
iSR

Hkg
hkg
HKG

Nzl
nzl
NZL

fin
SWE Nor swe
Fin
nor NOR FIN
SWE
CHE

Usa
usa
USA
AUS
Can Aus
aus CAN

jpn
irl
IRL

Irl
Jpn

Chl
chl

Kor KOR
kor
ARG

MEX Arg
arg

Ven

ven

IDN

Chn
Idn
CHN chn EGY Tur

Figure 4.11

bra

tur

PHL BOL Zaf

Bol

ind

Ind

zaf
ZAF
KEN

Egy
egy

nga
NGA

Ken
ken

pak Pak
PAK
Cmr
Uga
Nga CMR uga
cmh
UGA

bol
Jor

IND

bgd
BGD Bgd

Bra

Ecu
ecu
Phl
phl

Ita

hun
Esp CZE HUN
POL
CZe
Prt cze
pol
Hun
Pol

Tha
TUR

COL

idn

Fra
deu

Dnk
dnk
DNK

BRA tha

ECU

Col
col

Bel
ESP ITA
Deu GRC
PRT
Grc ita
grc esp

Rus
rus
RUS

THA

MYS Mex mex

mys
VEN

prt

BEL
Gbr
JPN
FRA
NLD gbr DEU
GBR
bel fra

CHL

ISR

jor Mys
JOR

Aut
aut

can

TWN
Twn
twn

Che
che

NID
Nld
AUT

Clustering Countries Using SOMs

Figure 4.12 provides the component plane for literacy whereby darker
regions score worse on literacy. Figure 4.13 provides the component
plane for political rights whereby darker regions correspond to better
political rights. It can be seen that many of the European countries
score good on both literacy and political rights.
SOMs are a very handy tool for clustering high‐dimensional data
sets because of the visualization facilities. However, since there is no
real objective function to minimize, it is harder to compare various
SOM solutions against each other. Also, experimental evaluation and
expert interpretation are needed to decide on the optimal size of the
SOM. Unlike k‐means clustering, an SOM does not force the number
of clusters to be equal to the number of output neurons.

Using and Interpreting Clustering Solutions
In order to use a clustering scheme, one can assign new observations
to the cluster for which the centroid is closest (e.g., in Euclidean or

Figure 4.12

Component Plane for Literacy

Figure 4.13 Component Plane for Political Rights

▸ A N A LY T I C S I N A BI G D ATA WO RL D

104

Manhattan sense). To facilitate the interpretation of a clustering solution, one could do the following:
■

Compare cluster averages with population averages for all variables using histograms, for example.

■

Build a decision tree with the cluster ID as the target and the
clustering variables as the inputs (can also be used to assign new
observations to clusters).

It is also important to check cluster stability by running different
clustering techniques on different samples with different parameter
settings and check the robustness of the solution.

NOTES
1. R. Agrawal, T. Imielinski, and A. Swami, “Mining Association Rules between Sets of
Items in Massive Databases,” in Proceedings of the ACM SIGMOD International Conference on Management of Data (Washington, DC, ACM, 1993).
2. R. Srikant and R. Agrawal, “Mining Generalized Association Rules,” in Proceedings of
the 1995 International Conference on Very Large Data Bases (Zurich, 1995).
3. T. Kohonen, “Self‐Organized Formation of Topologically Correct Feature Maps,” Biological Cybernetics 43 (1982): 59–69; J. Huysmans et al., “Using Self Organizing Maps
for Credit Scoring,” Special issue on Intelligent Information Systems for Financial
Engineering, Expert Systems With Applications, 30, no. 3 (2006): 479–487; A. Seret
et al., “A New SOM‐Based Method for Profile Generation: Theory and an Application in Direct Marketing,” European Journal of Operational Research 220, no. 1 (2012):
199–209.

C H A P T E R

Survival Analysis

urvival analysis is a set of statistical techniques focusing on the
occurrence and timing of events.1 As the name suggests, it originates from a medical context where it was used to study survival
times of patients that had received certain treatments. In fact, many
classification analytics problems we have discussed before also have a
time aspect included, which can be analyzed using survival analysis
techniques. Some examples are:2
■

Predict when customers churn

■

Predict when customers make their next purchase

■

Predict when customers default

■

Predict when customers pay off their loan early

■

Predict when customer will visit a website next

Two typical problems complicate the usage of classical statistical
techniques such as linear regression. A first key problem is censoring.
Censoring refers to the fact that the target time variable is not always
known because not all customers may have undergone the event yet at
the time of the analysis. Consider, for example, the example depicted
in Figure 5.1. At time T, Laura and John have not churned yet and thus
have no value for the target time indicator. The only information available is that they will churn at some later date after T. Note also that
Sophie is censored at the time she moved to Australia. In fact, these are
all examples of right censoring. An observation on a variable T is right
censored if all you know about T is that it is greater than some value c.
105

106

▸ A N A LY T I C S I N A BI G DATA WO RL D

Bart

Churn

Laura
Victor
Sophie

Churn
Churn
Moves to
Australia
Churn

John
T

Time

Example of Right Censoring for Churn Prediction

Likewise, an observation on a variable T is left censored if all you know
about T is that it is smaller than some value c. An example here could
be a study investigating smoking behavior and some participants at
age 18 already began smoking but can no longer remember the exact
date. Interval censoring means the only information available on T is
that it belongs to some interval a < T < b. Returning to the previous
smoking example, one could be more precise and say 14 < T < 18. Censoring occurs because many databases only contain current or rather
recent customers for whom the behavior has not yet been completely
observed, or because of database errors when, for example, the event
dates are missing. Using classical statistical analysis techniques such as
linear regression, the censored observations would have to be left out
from the analysis, since they have no value for the target time variable. However, with survival analysis, the partial information available
for the censored observations giving either a lower and/or an upper
bound on the timing of the event will be included in the estimation.
Time‐varying covariates are variables that change value during the
course of the study. Examples are account balance, income, and credit
scores. Survival analysis techniques will be able to accommodate this
in the model formulation, as will be discussed in what follows.

SURVIVAL ANALYSIS MEASUREMENTS
A first important concept is the event time distribution defined as a
continuous probability distribution, as follows:
f (t) = lim

Δt →0

P(t ≤ T < t + ΔT )
Δt

S U R V I V A L A N A LY S I S ◂

107

The corresponding cumulative event time distribution is then
defined as follows:
t

F(t) = P(T ≤ t) =

∫ f (u)du
0

Closely related is the survival function:
∞

S(t) = 1 − F(t) = P(T > t) =

∫ f (u)du
t

S(tt) is a monotonically decreasing function with S(0) = 1 and S(∞) = 0.
The following relationships hold:
f (t ) =

dF(t )
dS(t )
=−
dt
dt

Figure 5.2 provides an example of a discrete event time distribution, with the corresponding cumulative event time and survival
distribution depicted in Figure 5.3.
Another important measure in survival analysis is the hazard function, defined as follows:
h(t) = lim

Δt →0

P(t ≤ T < t + ΔT | T ≥ t)
Δt

The hazard function tries to quantify the instantaneous risk that
an event will occur at time t, given that the individual has survived up
to time t. Hence, it tries to measure the risk of the event occurring at
time point t. The hazard function is closely related to the event time
60%
Frequency

50%
40%
30%
20%
10%
0%
1

5
Month

Figure 5.2 Example of a Discrete Event Time Distribution

▸ A N A LY T I C S I N A BI G DATA WO RL D

Frequency

108

100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%

Survival function
Cumulative distribution

Month
Cumulative Distribution and Survival Function for the Event Time
Distribution in Figure 5.2

distribution up to the conditioning on T ≥ t . That is why it is often also
referred to as a conditional density.
Figure 5.4 provides some examples of hazard shapes, as follows:
■

Constant hazard, whereby the risk remains the same at all times.

■

Increasing hazard, reflecting an aging effect.

■

Decreasing hazard, reflecting a curing effect.

■

Convex bathtub shape, which is typically the case when studying human mortality, since mortality declines after birth and
infancy, remains low for a while, and increases with elder years.
It is also a property of some mechanical systems to either fail
soon after operation, or much later, as the system ages.

The probability density function f(t), survivor function S(t), and the
hazard function h(t) are mathematically equivalent ways of describing
a continuous probability distribution with the following relationships:
h(t ) =

f (t )
S(t )

h(t ) = −

dlogS(t )
dt

⎛ t
⎞
S(t ) = exp ⎜ − ∫ h(u)du⎟
⎝ 0
⎠

S U R V I V A L A N A LY S I S ◂

109

Figure 5.4 Example Hazard Shapes

KAPLAN MEIER ANALYSIS
A first type of survival analysis is Kaplan Meier (KM) analysis, which
is also known as the product limit estimator or nonparametric maximum likelihood estimator for S(t). If no censoring is available in the data
set, the KM estimator for S(t) is just the sample proportion with event
times greater than t. If censoring is present, the KM estimator starts with
ordering the event times in ascending order t1 < t2 < … < tk. At each time
tj, there are nj individuals who are at risk of the event. At risk means that
they have not undergone the event, nor have they been censored prior
to tj. Let dj be the number of individuals who die (e.g., churn, respond,
default) at tj. The KM estimator is then defined as follows:
Sˆ(t ) =

⎛

dj ⎞

⎛

d ⎞

∏ ⎜⎝ 1 − n j ⎟⎠ = Sˆ(t − 1) i ⎜⎝ 1 − ntt ⎟⎠ = Sˆ(t − 1) i (1 − h(t))

j :t j ≤ t

for t1 ≤ t ≤ tk. The intuition of the KM estimator is very straightforward
because it basically states that in order to survive time t, one must
survive time t − 1 and cannot die during time t.
Figure 5.5 gives an example of Kaplan Meier analysis for churn
prediction.

110

▸ AN A LY T I C S IN A BI G D ATA WO RL D

Customer
C1
C2
C3
C4
C5
C6
C7
C8
C9
C10

Time of Churn or Churn or
Censoring
Censored
6
3
12
15
18
12
3
12
9
15
Time

Churn
Censored
Churn
Censored
Censored
Churn
Churn
Churn
Censored
Churn
Customers at Risk
Customers
Customers Censored
at t (nt)
Churned at t (dt)
at t

S(t)

0.9

0.9* 7/8 = 0.79

0.79* 7/7 = 0.79

0.79* 3/6 = 0.39

0.39* 2/3 = 0.26

0.26* 1/1 = 0.26

Figure 5.5 Kaplan Meier Example

If there are many unique event times, the KM estimator can be
adjusted by using the life table (also known as actuarial)
l method to
group event times into intervals as follows:
Sˆ(t ) =

⎡

∏ ⎢1 − n

j:t j ≤t ⎣

⎤
⎥
− c j /2 ⎦
dj

which basically assumes that censoring occurs uniformly across
the time interval, such that the average number at risk equals (nj +
(nj − cj))/2 or nj − cj/2.
Kaplan Meier analysis can also be extended with hypothesis testing to see whether the survival curves of different groups (e.g., men
versus women, employed versus unemployed) are statistically different. Popular test statistics here are the log‐rank test (also known as the
Mantel‐Haenzel testt), the Wilcoxon test, and the likelihood ratio statistic,
which are all readily available in any commercial analytics software.
KM analysis is a good way to start doing some exploratory survival
analysis. However, it would be nice to be able to also build predictive
survival analysis models that take into account customer heterogeneity
by including predictive variables or covariates.

S U R V I V A L A N A LY S I S ◂

111

PARAMETRIC SURVIVAL ANALYSIS
As the name suggests, parametric survival analysis models assume a
parametric shape for the event time distribution. A first popular choice
is an exponential distribution, defined as follows:
f (t) = λe −λt
Using the relationships defined earlier, the survival function then
becomes:
S(t) = e −λt
and the hazard rate
h(t ) =

f (t )
=λ
S(t )

It is worth noting that the hazard rate is independent of time such
that the risk always remains the same. This is often referred to as the
memoryless property of an exponential distribution. Figure 5.6 shows an
example of an exponential event time distribution together with its
cumulative distribution and hazard function.
When taking into account covariates, the model becomes:
log(h(t , x i )) = μ + β1 x i1 + β2 x i 2 + β N x iN

1
0.9
0.8
0.7
0.6

Hazard

0.5

S(t)

0.4

f(t)

0.3
0.2
0.1
0

Exponential Event Time Distribution, with Cumulative Distribution and
Hazard Function

112

▸ A N A LY T I C S I N A BI G DATA WO RL D

Note that the logarithmic transform is used here to make sure that
the hazard rate is always positive.
The Weibull distribution is another popular choice for a parametric
survival analysis model. It is defined as follows:
f (t ) = κρ(ρt )κ−1exp[−(ρt )κ ]
The survival function then becomes:
S(t ) = exp[−(ρt )κ ]
and the hazard rate
f (t )
= κρ(ρt )κ−1
S(t )

h(t ) =

Note that in this case the hazard rate does depend on time and can
be either increasing or decreasing (depending upon κ and ρ).
When including covariates, the model becomes:
log ( h(t , x i )) = μ + α log(t ) + β1 x i1 + β2 x i 2 + β N x iN
Other popular choices for the event time distribution are the
gamma, log‐logistic, and log‐normal distribution.3
Parametric survival analysis models are typically estimated using
maximum likelihood procedures. In case of no censored observations,
the likelihood function becomes:
n

L = ∏ f (t i )
i =1

When censoring is present, the likelihood function becomes:
n

L = ∏ f ( t i ) S(t i )1−δ
δi

i =1

δ i equals 0 if observation i is censored, and 1 if the observation dies at time ti. It is important to note here that the censored observations do enter the likelihood function and, as such, have an impact
on the estimates. For example, for the exponential distribution, the
likelihood function becomes:
n

L = ∏[λe −λt ]δ [e −λt ]1−δ
i

i =1

S U R V I V A L A N A LY S I S ◂

113

This maximum likelihood function is then typically optimized by
further taking the logarithm and then using a Newton Raphson optimization procedure.
A key question concerns the appropriate event time distribution
for a given set of survival data. This question can be answered both in
a graphical and a statistical way.
In order to solve it graphically, we can start from the following
relationships:
h(t ) = −

dlogS(t )
dt

or
t

− log(S(t)) = ∫ h(u) du
0

Because of this relationship, the log survivor function is commonly
referred to as the cumulative hazard function, denoted as Λ(t ). It can be
interpreted as the sum of the risks that are faced when going from time
0 to time t. If the survival times are exponentially distributed, then
the hazard is constant, h(t ) = λ , hence Λ(t ) = λt and a plot of –log(S(t))
versus t should yield a straight line through the origin at 0. Similarly,
it can be shown that if the survival times are Weibull distributed,
then a plot of log(−log(S(tt)) versus log(t) should yield a straight line
(not through the origin) with a slope of κ. These plots can typically
be asked for in any commercial analytics software implementing survival analysis. Note, however, that this graphical method is not a very
precise method because the lines will never be perfectly linear or go
through the origin.
A more precise method for testing the appropriate event time
distribution is a likelihood ratio test. In fact, the likelihood ratio
test can be used to compare models if one model is a special case of
another (nested models). Consider the following generalized gamma
distribution:
β ⎛ t ⎞ kβ−1 −⎛⎜⎝ θ ⎞⎟⎠
e
f (t) =
⎜ ⎟
Γ(t)θ ⎝ θ ⎠
t

▸ A N A LY T I C S I N A BI G DATA WO RL D

114

1
Let’s now use the following shortcut notations: σ =
and
β k
1
, then the Weibull, exponential, standard gamma, and log‐
δ=
k
normal model are all special versions of the generalized gamma model,
as follows:
■

σ = δ: standard gamma

■

δ = 1: Weibull

■

σ = δ = 1: exponential

■

δ = 0: log‐normal

Let Lfull now be the likelihood of the full model (e.g., generalized
gamma) and Lred be the likelihood of the reduced (specialized) model
(e.g., exponential). The likelihood ratio test statistic then becomes:
⎛L ⎞
−2log ⎜ red ⎟ ∼ χ2(k)
⎝ L full ⎠
whereby the degrees of freedom k depends on the number of parameters that need to be set to go from the full model to the reduced model.
In other words, it is set as follows:
■

Exponential versus Weibull: one degree of freedom

■

Exponential versus standard gamma: one degree of freedom

■

Exponential versus generalized gamma: two degrees of freedom

■

Weibull versus generalized gamma: one degree of freedom

■

Log‐normal versus generalized gamma: one degree of freedom

■

Standard gamma versus generalized gamma: one degree of
freedom

The χ ‐test statistic can then be calculated together with the corresponding p‐value and a decision can be made about what is the most
appropriate event time distribution.
2

PROPORTIONAL HAZARDS REGRESSION
The proportional hazards model is formulated as follows:
h(t , x i ) = h0(t)exp(β1x i1 + β2 x i 2 + … + β N x iN )

S U R V I V A L A N A LY S I S ◂

115

so the hazard of an individual i with characteristics x i at time t is the
product of a baseline hazard function h0(t ) and a linear function of a set
of fixed covariates, which is exponentiated. In fact, h0(t) can be considered as the hazard for an individual with all covariates equal to 0. Note
that if a variable j increases with one unit and all other variables keep
their values (ceteris paribus), then the hazards for all t increase with
exp(β j ), which is called the hazard ratio (HR). If β j > 0 then HR > 1,
β j < 0 then HR < 1; β j = 0 then HR = 1. This is one of the most popular
models for doing survival analysis.
The name proportional hazards stems from the fact that the hazard
of any individual is a fixed proportion of the hazard of any other
individual.
hi (t )
= exp(β1( x i1 − x j1) + β1( x i 2 − x j 2) + + βn( x iN − x jN )).
h j (t )
Hence, the subjects most at risk at any one time remain the subjects most at risk at any one other time (see also Figure 5.7).
Taking logarithms from the original proportional hazards model
gives:
log h(t , x i ) = α(t ) + β1 x i1 + β2 x i 2 + + β N x iN
Note that if one chooses α(t ) = α , one gets the exponential model,
whereas if α(t ) = α log(t ), the Weibull model is obtained. A nice property of the proportional hazards model is that, using the idea of partial
likelihood, the βs can be estimated without having to explicitly specify
the baseline hazard function h0(t ).4 This is useful if one is only interested in analyzing the impact of the covariates on the hazard rates and/
or survival probabilities. However, if one wants to make predictions
Log h(t)
Subject i
Subject j

Figure 5.7 The Proportional Hazards Model

116

▸ A N A LY T I C S I N A BI G DATA WO RL D

with the proportional hazards model, the baseline hazard needs to be
explicitly specified.
The survival function that comes with the proportional hazards
model looks like this:
⎤
⎡ t
S(t , x i ) = exp ⎢ − ∫ h0(u)exp(β1 x i1 + β2 x i 2 + + β N x iN )du ⎥ , or
⎥⎦
⎢⎣ 0
S(t , x i ) = S0(t )exp(β x +β x ++β x ) , with
1 i1

⎛ t
⎞
S0(t ) = exp ⎜ − ∫ h0(u)du⎟
⎝ 0
⎠
S0(tt) is referred to as the baseline survivor function, that is, the survivor
function for an individual whose covariates are all 0. Note that if a
variable j increases with one unit (ceteris paribus), the survival probabilities are raised to the power exp(β j), which is the hazard ratio (HR).

EXTENSIONS OF SURVIVAL ANALYSIS MODELS
A first extension of the models we previously discussed is the inclusion of time‐varying covariates. These are variables that change value
throughout the course of the study. The model then becomes:
h(t , x i ) = h0(t )exp(β1 x i1(t) + β2 x i 2(t) + + β N x iN (t))
Note that the proportional hazards assumption here no longer
holds because the time‐varying covariates may change at different
rates for different subjects, so the ratios of their hazards will not remain
constant. One could also let the β parameters vary in time, as follows:
h(t , x i ) = h0(t)exp(β1(t)x i1(t) + β2(t)x i 2(t) + + β N (t)x iN (t))
The partial likelihood estimation method referred to earlier can
easily be extended to accommodate these changes in the model formulation, such that the coefficients can also be estimated without
explicitly specifying the baseline hazard h0(t ).
Another extension is the idea of competing risks.5 Often, an
observation can experience any of k competing events. In medicine,
customers may die because of cancer or ageing. In a bank setting, a

S U R V I V A L A N A LY S I S ◂

117

customer can default, pay off early, or churn at a given time. As long
as a customer has not undergone any of the events, he or she remains
at risk for any event. Once a customer has undergone the event, he or
she is no longer included in the population at risk for any of the other
risk groups, hence he or she becomes censored for the other risks.
Although the ideas of time‐varying covariates and competing risks
seem attractive at first sight, the number of successful business applications of both remains very limited, due to the extra complexity introduced in the model(s).

EVALUATING SURVIVAL ANALYSIS MODELS
A survival analysis model can be evaluated by first considering the statistical significance of both the model as a whole and the individual
covariates. (Remember: Significant covariates have low p‐values.) One
could also predict the time of the event when the survival curve S(t)
drops below 0,50 and compare this with the real event time. Another
option is to take a snapshot of the survival probabilities at a specific time
t (e.g., 12 months), compare this with the event time indicator, and calculate the corresponding ROC curve and its area beneath. The AUC will
then indicate how well the model ranks the observations for a specific
timestamp t. Finally, one could also evaluate the interpretability of the
survival analysis model by using univariate sign checks on the covariates and seeing whether they correspond to business expert knowledge.
The survival analysis models we have discussed in this chapter are
classical statistical models. Hence, some important drawbacks are that
the functional relationship remains linear or some mild extension
thereof, interaction and nonlinear terms have to be specified ad hoc,
extreme hazards may occur for outlying observations, and there is the
assumption of proportional hazards that may not always be the case.
Other methods have been described in the literature to tackle these
shortcomings, based on, for example, splines and neural networks.6

NOTES
1. P. D. Allison, Survival Analysis Using the SAS System (SAS Institute Inc., Cary, NC, US,
1995); D. R. Cox, “Regression Models and Life Tables,” Journal of the Royal Statistical
Society, series B (1972); D. R. Cox and D. Oakes, Analysis of Survival Data (Chapman

118

3.
4.

5.
6.

▸ A N A LY T I C S I N A BI G DATA WO RL D

and Hall, 1984); D. Kalbfleisch and R. L. Prentice, The Statistical Analysis of Failure
Time Data (New York: Wiley, 2003).
J. Banasik, J. N. Crook, and L. C. Thomas, “Not If but When Borrowers Will Default,”
Journal of the Operational Research Society 50, no. 12 (1999): 1185–1190; L. C. Thomas
and M. Stepanova, “Survival Analysis Methods for Personal Loan Data,” Operations
Research 50 (2002): 277–289.
P. D. Allison, Survival Analysis using the SAS System (SAS Institute Inc., Cary, NC, US,
1995).
P. D. Allison, Survival Analysis Using the SAS System (SAS Institute Inc., Cary, NC,
US,1995); D. R. Cox, “Regression Models and Life Tables,” Journal of the Royal
Statistical Society, series B (1972); D. R. Cox and D. Oakes, Analysis of Survival Data
(Chapman and Hall, 1984); D. Kalbfleisch and R. L. Prentice, The Statistical Analysis of
Failure Time Data (New York: Wiley, 2003).
M. J. Crowder, Classical Competing Risks (London: Chapman and Hall, 2001).
B. Baesens et al., “Neural Network Survival Analysis for Personal Loan Data.” Special issue, Journal of the Operational Research Society 59, no. 9 (2005): 1089–1098.

C H A P T E R

Social Network
Analytics

any types of social networks exist. The most popular are
undoubtedly Facebook, Twitter, Google+, and LinkedIn. However, social networks are more than that. It could be any set of
nodes (also referred to as vertices) connected by edges in a particular
business setting. Examples of social networks could be:
■

Web pages connected by hyperlinks

■

Email traffic between people

■

Research papers connected by citations

■

Telephone calls between customers of a telco provider

■

Banks connected by liquidity dependencies

■

Spread of illness between patients

These examples clearly illustrate that social network analytics can
be applied in a wide variety of different settings.

SOCIAL NETWORK DEFINITIONS
A social network consists of both nodes (vertices) and edges. Both
need to be clearly defined at the outset of the analysis. A node (vertex)
could be defined as a customer (private/professional), household/
family, patient, doctor, paper, author, terrorist, web page, and so forth.
An edge can be defined as a friend relationship, a call, transmission
119

120

▸ A N A LY T I C S I N A BI G DATA WO RL D

Figure 6.1 Example Sociogram

of a disease, reference, and so on. Note that the edges can also be
weighted based on interaction frequency, importance of information
exchange, intimacy, and emotional intensity. For example, in a churn
prediction setting, the edge can be weighted according to the time two
customers called each other during a specific period. Social networks
can be represented as a sociogram. This is illustrated in Figure 6.1,
whereby the color of the nodes corresponds to a specific status (e.g.,
churner or nonchurner).
Sociograms are good for small‐scale networks. For larger‐scale
networks, the network will typically be represented as a matrix, as
illustrated in Table 6.1. These matrices will be symmetrical and typically very sparse (with lots of zeros). The matrix can also contain the
weights in case of weighted connections.

S O C I A L N E T W O R K A N A LY T I C S

Table 6.1

◂

121

Matrix Representation of a Social Network

—

SOCIAL NETWORK METRICS
A social network can be characterized by various social network
metrics. The most important centrality measures are depicted in
Table 6.2. Assume a network with g nodes ni, i = 1, …, g. gjk represents the number of geodesics from node j to node k, whereas gjk(ni)
represents the number of geodesics from node j to node k passing
through node ni. The formulas each time calculate the metric for
node ni.
These metrics can now be illustrated with the well‐known Kite
network depicted in Figure 6.2.
Table 6.3 reports the centrality measures for the Kite network.
Based on degree, Diane has the most connections. She works as a
Table 6.2

Network Centrality Measures

Geodesic

Shortest path between two
nodes in the network

Degree

Number of connections of a
node (in‐ versus out‐degree if
the connections are directed)

Closeness

The average distance of a
node to all other nodes in
the network (reciprocal of
farness)

Betweenness

Counts the number of times
a node or connection lies on
the shortest path between
any two nodes in the network

Graph theoretic center

The node with the smallest
maximum distance to all
other nodes in the network

⎡ ∑ g d(ni n j ) ⎤
⎥
⎢ j =1
⎥
⎢
g
⎦
⎣

∑
j 5 and n(1 − Pˆ ) > 5, the binomial distribution can
be approximated by a normal distribution as N( nP̂ , nPˆ (1 − Pˆ )). Hence,
one obtains:
⎛
k * − nPˆ ⎞⎟
P ⎜⎜ z ≤
⎟ = α,
nPˆ (1 − Pˆ ) ⎠
⎝
with z a standard normally distributed variable. The critical value, k*,
can then be obtained as follows:
k * = nPˆ + N −1(α) nPˆ (1 − Pˆ )

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

◂

141

with N −1(α) the inverse cumulative standard normal distribution. In
terms of a critical event rate, p*, one then has:
p* = Pˆ + N −1(α)

Pˆ (1 − Pˆ )
n

H0 can then be rejected at significance level α, if the observed
event rate is higher than p*. Remember that the binomial test
assumes that all observations are independent. If the observations
are correlated, then the binomial test has a higher probability to
erroneously reject H0 (type I error), so that’s why it is often used
as an early warning system. It can be coded using traffic lights, as
follows:
■

Green (normal font): no statistical difference at 90 percent

■

Yellow (italics): statistical difference at 90 percent but not at
95 percent

■

Orange (bold face): statistical difference at 95 percent but not
at 99 percent

■

Red (bold face and underlined): statistical difference at
99 percent

Table 7.7 shows an example of using the binomial test for backtesting calibrated probabilities of default (PDs) against observed default
rates (DRs). It can be seen that from 2001 onwards, the calibration is
no longer satisfactory.
The Hosmer‐Lemeshow test is a closely related test that will test
calibrated versus observed event rates across multiple segments/pools
simultaneously. It also assumes independence of the events, and the
test statistic is defined as follows:
k

(ni Pˆi − θi )2
n Pˆ (1 − Pˆi )
i =1 i i

χ2 (k) = ∑

whereby ni is the number of observations in pool i, Pˆi is the estimated
probability of the event for pool i, and θi is the number of observed
events. The test statistic follows a chi‐squared distribution with k degrees
of freedom. It can be coded using traffic lights in a similar way as for the
binomial test.

142

0.00%

0.28%

0.27%

1.26%

0.26%

1997

1998

1999

2000

2001

2002

0.00%

1995

0.00%

1994

1996

0.00%

1993

0.17%

0.72%

0.27%

0.00%

0.31%

0.00%

Baa2

0.17%

0.26%

Baa1

Baa2

Baa1

0.42%

1.78%

0.00%

0.97%

0.34%

0.00%

Baa3

0.42%

Baa3

0.53%

1.58%

0.51%

0.94%

0.47%

0.00%

0.83%

Ba1

0.53%

Ba1

0.54%

1.41%

1.38%

0.63%

0.00%

0.62%

0.00%

Ba2

0.54%

Ba2

The Binomial Test for Backtesting PDs versus DRs

Table 7.7

Ba3

1.36%

1.58%

2.93%

1.04%

2.00%

1.12%

0.47%

0.00%

1.76%

0.59%

0.76%

Ba3

1.36%

2.46%

2.00%

3.19%

3.24%

3.28%

2.11%

0.00%

1.17%

4.35%

1.88%

3.24%

2.46%

5.76%

6.81%

11.07%

4.10%

6.91%

7.55%

1.54%

0.00%

6.42%

3.75%

5.04%

5.76%

8.76%

6.86%

16.38%

10.88%

9.63%

5.52%

7.22%

3.28%

4.06%

7.95%

11.29%

8.76%

Caa‐C

20.9%

29.45%

34.45%

19.65%

20.44%

15.09%

14.67%

13.99%

11.57%

5.13%

28.57%

Caa‐C

20.89%

3.05%

3.70%

5.48%

3.01%

3.35%

2.83%

1.41%

0.78%

2.51%

1.88%

3.24%

3.05%

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

◂

143

Backtesting Regression Models
In backtesting regression models, one can also make a distinction
between model ranking and model calibration. When predicting
CLV, one might especially be interested in model ranking, since it is
typically hard to accurately quantify CLV. However, in the majority
of the cases, the aim is model calibration. For ranking, one could first
consider a system stability index (SSI), as discussed before, applied
to the categorized output. Also t‐tests and/or histograms can be used
here. For ranking, one could create a scatter plot and summarize it
into a Pearson correlation coefficient (see Chapter 3). For calibration, one can calculate the R‐squared, mean squared error (MSE),
or mean absolute deviation (MAD) as also discussed in Chapter 3.
Table 7.8 gives an example of a table that can be used to monitor
the MSE.

Backtesting Clustering Models
When backtesting clustering models, one can first check the data
stability by comparing the number of observations per cluster during model design with the number observed now and calculate a
system stability index (SSI) across all clusters. One can also measure
how the distance/proximity measures have changed on new observations by creating histograms of distances per cluster and compare
the histograms of the model design data with those of new data. The

Table 7.8 Monitoring Model Calibration Using MSE

MSE
MSE model
MSE year t
MSE year t + 1
MSE year t + 2
…
Average MSE period 1
Average MSE period 2

Number of
Observations

Number of
Events

Traffic Light

▸ A N A LY T I C S I N A BI G DATA WO RL D

144

distances can then be statistically tested using, for example, a t‐test.
One can also statistically compare the intracluster similarity with the
intercluster similarity using an F‐test to see whether reclustering is
needed.

Developing a Backtesting Framework
In order to setup a backtesting framework, one needs to decide on the
following:
■

Diagnose backtesting needs

■

Work out backtesting activities

■

Design timetable for backtesting activities

■

Specify tests and analyses to be performed

■

Define actions to be taken in response to findings

■

Identify why/what/who/how/when

All of the above should be described in a backtesting policy.
Figure 7.2 presents an example of a digital dashboard application that
could be developed for backtesting classification models. Note also that
qualitative checks are included that are based on a judgment made by
one or more business experts. These subjective evaluations are considered to be very important.
Once a backtesting framework has been developed, it should be
complemented with an action plan. This plan will specify what to do
in response to what finding of the backtesting exercise. Figure 7.3
gives an example of this. If the model calibration is okay, one can
continue to use the model. If not, one needs to verify the model
discrimination or ranking. If this is okay, then the solution might be
to simply recalibrate the probabilities upward or downward using
a scaling factor. If not, the next step is to check the data stability. If
the data stability is still okay, one may consider tweaking the model.
Note that this is, however, not that straightforward and will often
boil down to reestimating the model (as is the case when the data
stability is not okay).

◂

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

Green

Qualitative

Not significant
at 95% level

Significant at
95% but not at
99% level

Significant at 99%
level

Hosmer-Lemeshow

Not significant
at 95% level

Significant at
95% but not at
99% level

Significant at 99%
level

Vasicek

Not significant
at 95% level

Significant at
95% but not at
99% level

Significant at 99%
level

Normal

Not significant
at 95% level

Significant at
95% but not at
99% level

Significant at 99%
level

Portfolio distribution

Minor shift

Moderate shift

Major shift

Difference

Correct

Overestimation

Underestimation

Portfolio stability

Minor
migrations

Moderate
migrations

Major migrations

AR difference with
reference model

< 5%

Between 5%
and 10%

> 10%

< 2.5%

Between 2.5%
and 5%

> 5%

Model significance

p-value < 0.01

p-value between p-value > 0.10
0.01 and 0.10

Preprocessing
(missing values,
outliers)

Considered

Partially
considered

Coefficient signs

All as expected Minor
exceptions

Major exceptions

Number of overrides

Minor

Moderate

Major

Documentation

Sufficient

Minor issues

Major issues

Green

Quantitative AUC difference with
reference model

Level 1:
Discrimination

Qualitative

Green

Quantitative
Level 0: Data

Qualitative

Figure 7.2

Red

Binomial

Quantitative
Level 2:
Calibration

Yellow

145

Yellow

Red

Ignored

Red

SSI (current versus
training sample)

SSI < 0.10

0.10< SSI< 0.25 SSI > 0.25

SSI attribute level

SSI < 0.10

0.10< SSI< 0.25 SSI > 0.25

t-test attribute level

p-value > 0.10 p-value between p-value < 0.01
0.10 and 0.01

Characteristic
analysis

No change

Moderate
change

Major change

Attribute histogram

No shift

Moderate shift

Major shift

A Backtesting Digital Dashboard for Classification Models

146

▸ A N A LY T I C S I N A BI G DATA WO RL D

Model
calibration
Not Okay

Okay

Model
discrimination
Not Okay

Okay
Recalibrate
model

Data stability
Not Okay
Reestimate
model

Continue using
model

Okay
Tweak
model

Figure 7.3 Example Backtesting Action Plan

BENCHMARKING
The idea of benchmarking is to compare the output and performance
of the analytical model with a reference model or benchmark. This is
needed as an extra validity check to make sure that the current analytical is the optimal one to be used. The benchmark can be externally or
internally developed. A popular example of an external benchmark in
credit risk modeling could be the FICO score. This is a credit score that
ranges between 300 and 850 and is developed by Experian, Equifax,
and Transunion in the United States. It is often used as a benchmark to
compare application and/or behavioral credit scoring models. A closely
related score is the Vantage score, also available in the United States.
Credit rating agencies (e.g., Moody’s, S&P, and Fitch) could also be
considered as benchmarking partners. These agencies typically provide
information on credit ratings and default probabilities that are very
useful in a credit risk modeling context.
Note that although external benchmarking may seem appealing
at first sight, one should be aware of potential problems, for example,
unknown quality of the external benchmark, different underlying data
samples and/or methodologies, different target definitions, and legal
constraints. One should also be vigilant for cherry‐picking, whereby

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

◂

147

the external benchmark is selected so as to correspond as closely as
possible to the internal model.
The benchmark can also be internally developed, either statistically or expert based. For example, one could benchmark a logistic regression model against a neural network benchmark to see
whether there are any significant nonlinearities in the data. If it
turns out that this is indeed the case, then nonlinear transformations and/or interaction terms can be added to the logistic regression model to come as close as possible to the neural network performance. An expert-based benchmark is a qualitative model based
on expert experience and/or common sense. An example of this
could be an expert committee ranking a set of small‐ and medium‐
sized enterprises (SMEs) in terms of default risk by merely inspecting their balance sheet and financial statement information in an
expert‐based, subjective way.
When benchmarking, one commonly adopts a champion–
challenger approach. The current analytical model serves as the
champion and the benchmark as the challenger. The purpose of the
challenger is to find the weaknesses of the champion and to beat it.
Once the benchmark outperforms the champion, one could consider
making it the new champion, and the old champion then becomes the
new benchmark. The purpose of this approach is to continuously challenge the current model so as to continuously perfect it.
Popular agreement statistics for benchmarking are Spearman’s
rank order correlation, Kendall’s τ , and the Goodman‐Kruskal γ .
Spearman’s rank order correlation measures the degree to which a
monotonic relationship exists between the scores or ratings provided
by an internal scoring system and those from a benchmark. It starts by
assigning 1 to the lowest score, 2 to the second lowest score, and so
on. In case of tied scores, the average is taken. Spearman’s rank order
correlation is then computed as follows:
ρs = 1−

6∑

d2
i =1 i
2

n(n − 1)

whereby n is the number of observations and di the difference between
the scores. Spearman’s rank order correlation always ranges between
−1 (perfect disagreement) and +1 (perfect agreement).

148

▸ A N A LY T I C S I N A BI G DATA WO RL D

Kendall’s τ works by first calculating the concordant and discordant pairs of observations. Two observations are said to be concordant if the observation that has a higher score assigned by the internal
model also has a higher score assigned by the external model. If there
is disagreement in the scores, then the pair is said to be discordant.
Note that if the pair is neither concordant nor discordant, it is tied,
meaning the two observations have identical scores assigned by the
internal model, or by the benchmark, or by both. Kendall’s τ is then
calculated as follows:
τ=

A−B
,
1
n(n − 1)
2

whereby n is the number of observations, A the number of concordant
pairs, and B the number of discordant pairs. Note that the denominator gives all possible pairs for n observations. Kendall’s τ is 1 for perfect
agreement and −1 for perfect disagreement.
Kendall’s τ basically looks at all possible pairs of observations. The
Goodman‐Kruskal γ will only consider the untied pairs (i.e., either
concordant or discordant), as follows:
γ=

A−B
A+B

The Goodman‐Kruskal γ is +1 if there are no discordant pairs (perfect agreement), −1 if there are no concordant pairs (perfect disagreement), and 0 if there are equal numbers of concordant and discordant
pairs.
For example, consider the example in Table 7.9.
Spearman’s rank order correlation then becomes −0.025. The concordant pairs are as follows: C1,C3; C1,C4; C3,C4; C3,C5; and C4,C5.
The discordant pairs are: C1,C2; C2,C3; C3,C4; and C2,C5. The pair
C1,C5 is a tie. Kendall’s τ thus becomes: (5 − 4)/10 or 0.1 and the
Goodman‐Kruskal γ becomes (5 − 4)/(5 + 4) or 0.11.
In case of disagreement between the current analytical model
and the benchmark, it becomes interesting to see which is the best
model overall, or whether there are certain segments of observations where either the internal model or benchmark proves to be
superior. Based on this analysis, it can be decided to further perfect

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

Table 7.9

Customer

◂

149

Example for Calculating Agreement Statistics

Internal
Credit Score

FICO

Rank Internal
Score

Rank
External
Score

680

2.5

580

640

720

700

2.5

2.25

∑ di2

0.25

20.5

i =1

the current analytical model or simply proceed with the benchmark
as the new model.

DATA QUALITY
Corporate information systems consist of many databases linked
by real-time and batch data feeds.2 The databases are continuously
updated, as are the applications performing data exchange. This
dynamism has a negative impact on data quality (DQ), which is very
disadvantageous since DQ determines the value of the data to the analytical technique. Information and communication technology can be
used to further improve intrinsic value. Hence, high-quality data in
combination with good technology gives added value, whereas poorquality data with good technology is a big problem (remember the
garbage in, garbage out idea discussed in Chapter 2). Decisions made
based on bad data can create high losses for companies. Poor DQ
impacts organizations in many ways. At the operational level, it has
an impact on customer satisfaction, increases operational expenses,
and will lead to lowered employee job satisfaction. Similarly, at the
strategic level, it affects the quality of the (analytical) decision making process.3
Poor DQ are often experienced in everyday life. For example, the mistaken delivery of a letter is often associated with

150

▸ A N A LY T I C S I N A BI G DATA WO RL D

malfunctioning postal services. However, one of the causes of this
mistaken delivery can be an error in the address. Similarly, two
similar emails sent to the same recipient can be an indication of a
duplication error.
Moreover, the magnitude of DQ problems is continuously growing
following the exponential increase in the size of databases. This certainly qualifies DQ management as one of the most important business
challenges in today’s information‐based economy.
Data quality is often defined as “fitness for use,” which implies
the relative nature of the concept.4 Data with quality for one use
may not be appropriate for another use. For instance, the extent
to which data is required to be complete for accounting tasks may
not be required for analytical sales prediction tasks. More generally,
data that are of acceptable quality in one decision context may be
perceived to be of poor quality in another decision context, even
by the same individual. This is mainly because DQ is a multidimensional concept in which each dimension represents a single
aspect or construct of data items and also comprises both objective and subjective aspects. Some aspects are independent while
others depend on the type of task and/or experience of the data
user. Therefore, it is useful to define DQ in terms of its dimensions.
Table 7.10 shows the different DQ dimensions, their categories, and
definitions.5
Accuracy indicates whether the data stored are the correct values. For example if my birthdate is February 27, 1975, for a database that expects dates in USA format, 02/27/1975 is the correct
value. However, for a database that expects a European representation, the date 02/27/1975 is incorrect; instead 27/02/1975 is the
correct value.6
Another interesting dimension concerns the completeness of data.
The completeness dimension can be considered from different perspectives. Schema completeness refers to the extent to which entities
and attributes are not lacking from the schema. Column completeness
verifies whether a column of a table has missing values or not. Finally,
population completeness refers to the degree to which members of
the population are not present. As an example, population completeness is depicted in Table 7.11.7

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

Table 7.10

Category

Dimension

Definition: The Extent to Which . . .

Intrinsic

Accuracy

Data are regarded as correct

Believability

Data are accepted or regarded as true, real, and
credible

Objectivity

Data are unbiased and impartial

Reputation

Data are trusted or highly regarded in terms of their
source and content

Value‐added

Data are beneficial and provide advantages for
their use

Completeness

Data values are present

Relevancy

Data are applicable and useful for the task at hand

Appropriate
amount of data

The quantity or volume of available data is appropriate

Interpretability

Data are in appropriate language and unit and the
data definitions are clear

Ease of
understanding

Data are clear without ambiguity and easily
comprehended

Accessibility

Data are available or easily and quickly retrieved

Security

Access to data can be restricted and hence kept
secure

Representational

Accessibility

Table 7.11

Population Completeness

Name

Surname

Birth Date

Monica

Smith

04/10/1978

smith@abc.it

Yuki

Tusnoda

04/03/1968

Nulla

Rose

David

02/01/1937

Nullb

John

Edward

14/12/1955

Nullc

Not existing
Existing but unknown
c
Not known if existing
b

151

Data Quality Dimensions

Contextual

◂

▸ A N A LY T I C S I N A BI G DATA WO RL D

152

Tuple 2: Since the person represented by tuple 2 has no email
address, we can say that the tuple is complete.
Tuple 3: Since the person represented by tuple 3 has an email, but
its value is not known, we can say that the tuple is incomplete.
Tuple 4: If we do not know the person represented by tuple 4 has
an email or not, incompleteness may not be the case.
A next data quality dimension is believability, which is the extent
to which data is regarded as true and credible.
Accessibility refers to how easy the data can be located and
retrieved. From a decision making viewpoint, it is important that the
data can be accessed and delivered on time, so as to not needlessly
delay important decisions.
The dimension of consistency can be considered from various perspectives. A first example is the presence of redundant data (e.g. name,
address, …) in multiple data sources.
Another perspective is the consistency between related data attributes. For example, city name and zip code should be corresponding.
Another consistency perspective concerns the data format used. For
example, gender can be encoded as male/female, M/F, or 0/1. It is of
key importance that a uniform coding scheme is adopted so as to have
a consistent corporate wide data representation.
The timeliness dimension reflects how up‐to‐date the data is with
respect to the task for which it is used.
There are different DQ problem causes such as:
■

Multiple data sources: Multiple sources of the same data may
produce duplicates; a consistency problem.

■

Subjective judgment: Subjective judgment can create data bias;
objectivity problem.

■

Limited computing facilities: Lack of sufficient computing facilities limits data access; accessibility problem.

■

Size of data: Big data can give high response times; accessibility
problem.

Data quality can be improved through a total data quality management program. It consists of the four phases, as shown in Figure 7.4.8

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

• Define

153

• Assess
Identifying
important DQ
dimensions

Suggesting
improvement
actions

• Improve

◂

Assessing/
measuring DQ
level using the
important DQ
dimensions

Investigating
DQ problems
and analyzing
their major
causes

• Analyze

Figure 7.4 Data Quality Management Program

SOFTWARE
Different types of software can be used for doing analytics. A first distinction can be made between open source and commercial software.
Popular open source analytical workbenches are RapidMiner (formerly Yale), R, and Weka. Especially the latter has gained in importance and usage nowadays. In the commercial area, SAS, SPSS, Matlab, and Microsoft are well‐known vendors of analytical software.
Many of these vendors actually provide analytical solutions targeted
at specific industries (e.g., churn prediction in telco, fraud detection
in insurance) and hereby provide full coverage of the whole range of
analytical activities needed in the specific business setting.
Table 7.12 presents an overview of a KDnuggets poll asking about
software used in 2012 and 2013.
Based on Table 7.12, it can be concluded that RapidMiner and R,
two open source software solutions, are the most popular tools for
analytics. The distinction between open source and commercial is getting more and more difficult to make, since vendors like RapidMiner
have also started providing commercial versions of their software.

154

▸ AN A LY T I C S I N A BI G DATA WO RL D

Table 7.12

Results of KDnuggets Poll on Software Tools Used in Analytics in 2012 and 2013.

Legend: Bold: Free/Open Source tools
Normal case: Commercial tools

First bar: % users in 2013
Second bar: % users in 2012
39.2%
26.7%

Rapid‐I RapidMiner/RapidAnalytics free
edition (737), 30.9% alone
R (704), 6.5% alone
Excel (527), 0.9% alone
Weka/Pentaho (269), 5.6% alone
Python with any of numpy/scipy/
pandas/iPython packages (250), 0% alone
Rapid‐I RapidAnalytics/RapidMiner
Commercial Edition (225), 52.4% alone
SAS (202), 2.0% alone
MATLAB (186), 1.6% alone
StatSoft Statistica (170), 45.9% alone
IBM SPSS Statistics (164), 1.8% alone
Microsoft SQL Server (131), 1.5% alone
Tableau (118), 0% alone
IBM SPSS Modeler (114), 6.1% alone
KNIME free edition (110), 1.8% alone
SAS Enterprise Miner (110), 0% alone
Rattle (84), 0% alone
JMP (77), 7.8% alone
Orange (67), 13.4% alone
Other free analytics/data mining
software (64), 3.1% alone
Gnu Octave (54), 0% alone

37.4%
30.7%
28.0%
29.8%
14.3%
14.8%
13.3%
14.9%
12.0%
10.7%
12.7%
9.9%
10.0%
9.0%
14.0%
8.7%
7.8%
7.0%
5.0%
6.3%
4.4%
6.1%
6.8%
5.9%
21.8%
5.9%
5.8%
4.5%
4.1%
4.0%
3.6%
5.3%
3.4%
4.9%
2.9%

Source:: www.kdnuggets.com/polls/2013/analytics‐big‐data‐mining‐data‐science‐software.html.

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

◂

155

In addition, Microsoft Excel is still quite popular for doing analytics.
The average number of tools used was 3.

PRIVACY
The introduction of new technology, such as data analytics, brings
new privacy concerns. Privacy issues can arise in two ways.9 First, data
about individuals can be collected without these individuals being
aware of it. Second, people may be aware that data is collected about
them, but have no say in how the data is being used. Furthermore, it
is important to note that data analytics brings extra concerns regarding
privacy as compared to simple data collection and data retrieval from
databases.
Data analytics entails the use of massive amounts of data—possibly
combined from several sources, including the Internet—to mine for
hidden patterns. Hence, this technology allows for the discovery of
previously unknown relationships without the customer and company being able to anticipate this knowledge. Consider an example in
which three independent pieces of information about a certain customer lead to the customer being classified as a long‐term credit risk,
whereas the individual pieces of information would never have led to
this conclusion. It is exactly this kind of discovery of hidden patterns
that forms an additional threat to citizens’ privacy.
Moreover, previous work has shown that it is possible to construct
partial profiles of a person by crawling the web for small amounts of
nonsensitive information that is publicly available; often this information is voluntarily published by individuals through social networking
sites.10 Also, the individual pieces of nonsensitive information are not
harmful for one’s privacy. However, when all information is aggregated into a partial profile, this information can be used for criminal activities—such as stalking, kidnapping, identity theft, phishing,
scams—or for direct marketing by legitimate companies. It is again
important to note that this use of data is not anticipated by citizens,
hence privacy issues arise.
As illustrated by the previous examples, data analytics is more than
just data collection and information retrieval from vast databases. This
is recognized by the definition of data mining in several government

▸ A N A LY T I C S I N A BI G DATA WO RL D

156

reports. For example, the U.S. Government Accountability Office11
defined data mining as:
the application of database technology and techniques—
such as statistical analysis and modeling—to uncover hidden
patterns and subtle relationships in data and to infer rules
that allow for the prediction of future results.
In the August 2006 Survey of DHS Data Mining Activities, the
Department of Homeland Security (DHS) Office of the Inspector General (OIG) defined data mining as:12
the process of knowledge discovery, predictive modeling,
and analytics. Traditionally, this involves the discovery of
patterns and relationships from structured databases of
historical occurrences.
Several other definitions have been given, and generally these definitions imply the discovery of hidden patterns and the possibility for
predictions. Thus, simply summarizing historical data is not considered
data mining.
There are several regulations in place in order to protect an individual’s privacy. The Fair Information Practice Principles (FIPPs), which
were stated in a report of the U.S Department of Health, Education
and Welfare in 1973,13 have served as the main inspiration for the Privacy Act of 1974. In 1980, the Organization for Economic Cooperation
and Development (OECD) defined its “Guidelines on the Protection of
Privacy and Transborder Flows of Personal Data.” The following basic
principles are defined to safeguard privacy:14
■

Collection limitation principle: Data collection should be done
lawfully and with knowledge and consent of the data subject.

■

Data quality principle: The data should be relevant for the purpose it is collected for, accurate, complete, and up‐to‐date.

■

Purpose specification principle: The purposes of the data should
be specified before data collection and the use should be limited
to these purposes.

■

Use limitation principle: The data should not be used for other
purposes than specified, neither should it be disclosed to other

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

◂

157

parties without consent of the data subject (or by the authority
of law).
■

Safety safeguards principle: The data should be protected against
risks of loss, unauthorized access, use, modification, or disclosure of data.

■

Openness principle: There should be a policy of openness about
the developments, practices, and policies with respect to personal data.

■

Individual participation principle: An individual has the right
to obtain confirmation whether data exists about him or her, to
receive the data, to challenge data relating to him or her and to
have it erased or completed should the challenge be successful.

■

Accountability principle: A data controller can be held accountable for compliance with the above principles.

These guidelines are widely accepted, have been endorsed by the
U.S. Department of Commerce, and are the foundation of privacy laws
in many other countries (e.g., Australia, Belgium).
Given the increasing importance and awareness of privacy in
the context of analytics, more and more research is being conducted
on privacy preserving data mining algorithms. The parties that are
typically involved are: the record owner, the data publisher, and
the data recipient.15 A data publisher can be untrusted, in which case
the collection of records needs to be done anonymously. When the
data publisher is trusted, the record owners are willing to share their
information with the data publisher, but not necessarily with third
parties, and it is necessary to anonymize the data. This can be further
complicated when the data publisher is a nonexpert in the sense that
he or she is not aware that (and how) the data recipient can mine
the data.
The privacy of an individual is breached when an attacker can
learn anything extra about a record owner, possibly with the presence of any background knowledge from other sources.16 Consider an
example in which explicit identifiers are removed from a data set, but
there is a combination of a number of variables (e.g., age, zip code,
gender), which serves as a quasi‐identifier (QID). This means that it
is possible to link the record owner, by means of the QID, to a record

▸ A N A LY T I C S I N A BI G DATA WO RL D

158

Zip Code

Age

Gender

Zip Code

Age

Gender

83661

836**

83659

836**

83645

836**

Example of Generalization and Suppression to Anonymize Data

owner in another data set. To preserve privacy, there should be several
records in the data set with the same QID.
There are several classes of methods to anonymize data.17 A first
class of methods is generalization and suppression. These methods will
remove information from the quasi‐identifiers, until the records are
not individually identifiable, as illustrated in Figure 7.5.
Another group of techniques consists of anatomization and permutation, which groups and shuffles sensitive values within a QID
group, in order to remove the relationship between the QID and sensitive attributes. Perturbation methods change the data by adding noise,
swapping values, creating synthetic data, and so forth, based on the
statistical properties of the real data.18

MODEL DESIGN AND DOCUMENTATION
Some example questions that need to be answered from a model
design perspective are:
■

When was the model designed, and by who?

■

What is the perimeter of the model (e.g., counterparty types,
geographical region, industry sectors)?

■

What are the strengths and weaknesses of the model?

■

What data were used to build the model? How was the sample
constructed? What is the time horizon of the sample?

■

Is human judgment used, and how?

It is important that all of this is appropriately documented. In fact,
all steps of the model development and monitoring process should be
adequately documented. The documentation should be transparent

A N A LY T I C S : P U T T I N G I T A L L T O W O R K

◂

159

and comprehensive. It is advised to use document management systems with appropriate versioning facilities to keep track of the different versions of the documents. An ambitious goal here is to aim for a
documentation test, which verifies whether a newly hired analytical
team could use the existing documentation to continue development
or production of the existing analytical model(s).

CORPORATE GOVERNANCE
From a corporate governance perspective, it is also important that the
ownership of the analytical models is clearly claimed. A good practice
here is to develop model boards that take full responsibility of one
or more analytical models in terms of their functioning, interpretation, and follow-up. Also, it is of key importance that the board of
directors and senior management are involved in the implementation and monitoring processes of the analytical models developed.
Of course, one cannot expect them to know all underlying technical
details, but they should be responsible for sound governance of the
analytical models. Without appropriate management support, analytical models are doomed to fail. Hence, the board and senior management should have a general understanding of the analytical models.
They should demonstrate active involvement on an ongoing basis,
assign clear responsibilities, and put into place organizational procedures and policies that will allow the proper and sound implementation and monitoring of the analytical models. The outcome of the
monitoring and backtesting exercise must be communicated to senior
management and, if needed, accompanied by appropriate (strategic)
response. Given the strategic importance of analytical models nowadays, one sees a strong need to add a Chief Analytics Officer (CAO) to
the board of directors to oversee analytic model development, implementation, and monitoring.

NOTES
1. E. Lima, C. Mues, and B. Baesens, “Monitoring and Backtesting Churn Models,”
Expert Systems with Applications 38, no. 1 (2010): 975–982; G. Castermans et al., “An
Overview and Framework for PD Backtesting and Benchmarking.” Special issue,
Journal of the Operational Research Society 61 (2010): 359–373.

160

▸ A N A LY T I C S I N A BI G DATA WO RL D

2. H. T. Moges et al., “A Multidimensional Analysis of Data Quality for Credit Risk
Management: New Insights and Challenges,” Information and Management, 50:1,
43–58, 2014.
3. A. Maydanchik, Data Quality Assessmentt (Bradley Beach, NJ: Technics Publications,
2007), 20–21.
4. R. Y. Wang and D. M. Strong, “Beyond Accuracy: What Data Quality Means to Data
Consumers,” Journal of Management Information Systems 12, no. 4 (1996): 5–33.
5. Ibid.
6. Y. W. Lee, L. L. Pipino, J. D. Funk, and R. Y. Wang, Journey to Data Quality (London:
MIT Press, 2006), 67–108.
7. C. Batini and M. Scannapieco, Data Quality: Concepts, Methodologies and Techniques
(New York: Springer, 2006), 20–50.
8. G. Shankaranarayanan, M. Ziad, and R. Y. Wang, “Managing Data Quality in
Dynamic Decision Environments: An Information Product Approach,” Journal of
Database Managementt 14, no. 4 (2003): 14–32.
9. H. T. Tavani, “Informational Privacy, Data Mining, and the Internet,” Ethics and Information Technology 1, no. 2 (1999): 137–145.
10. M. Pontual et al., “The Privacy in the Time of the Internet: Secrecy vs Transparency,”
in Proceedings of the Second ACM Conference on Data and Application Security and Privacy
(ACM, 2012), ACM, New York, US, 133–140.
11. U.S. General Accounting Office (GAO), “Data Mining: Federal Efforts Cover a Wide
Range of Uses,” GAO‐04‐548 (May 2004), www.gao.gov/new.items/d04548.pdf.
12. U.S. Department of Homeland Security, Survey of DHS Data Mining Activities,
August 2006.
13. The report is entitled “Records, Computers and the Rights of Citizens.”
14. The documentation can be found at www.oecd.org/internet/ieconomy/oecdguideli
nesontheprotectionofprivacyandtransborderflowsofpersonaldata.htm.
15. B. Fung et al., “Privacy‐Preserving Data Publishing: A Survey of Recent Developments,” ACM Computing Surveys (CSUR) 42, no. 4 (2010): 14.
16. T. Dalenius, “Finding a Needle in a Haystack—or Identifying Anonymous Census
Record, Journal of Official Statistics 2, no. 3 (1986): 329–336.
17. B. Fung et al., “Privacy‐Preserving Data Publishing: A Survey of Recent Developments,” ACM Computing Surveys (CSUR) 42, no. 4 (2010): 14.
18. For more details about the specific techniques, the reader is referred to overview
papers such as J. Wang et al., “A Survey on Privacy Preserving Data Mining,” in First
International Workshop on Database Technology and Applications (IEEE, Washington, DC,
US, 2009), 111–114; and B. Fung et al., “Privacy‐Preserving Data Publishing: A Survey of Recent Developments,” ACM Computing Surveys (CSUR) 42, no. 4 (2010): 14.

C H A P T E R

Example
Applications

nalytics is hot and is being applied in a wide variety of settings.
Without claiming to be exhaustive, in this chapter, we will briefly
zoom into some key application areas. Some of them have been
around for quite some time, whereas others are more recent.

CREDIT RISK MODELING
The introduction of compliance guidelines such as Basel II/Basel III has
reinforced the interest in credit scorecards. Different types of analytical
models will be built in a credit risk setting.1 A first example are application scorecards. These are models that score credit applications based
on their creditworthiness. They are typically constructed by taking two
snapshots of information: application and credit bureau information
at loan origination and default status information 12 or 18 months
ahead. This is illustrated in Figure 8.1.
Table 8.1 provides an example of an application scorecard.
Logistic regression is a very popular application scorecard construction
technique due to its simplicity and good performance.2 For the scorecard
in Table 8.1, the following logistic regression with WOE coding was used:
P(Customer = good | age , employment , salary)
1
=
−(β0 +β1WOEage +β2WOEemployment +β3WOE salary )
1+ e
161

▸ A N A LY T I C S I N A BI G DATA WO RL D

162

Snapshot 1

Snapshot 2

t18
Good or Bad Payer?

Application
Data

Age
Income
Marital status
Savings amount
….

Credit
Bureau
Data

Bureau score
Delinquency history
Number of bureau checks
Number of outstanding credits
….

Figure 8.1 Constructing a Data Set for Application Scoring

Typically, the model will then be re‐expressed in terms of the log
odds, as follows:
⎛ P(Customer = good | age , employment , salary)⎞
log ⎜
⎝ P(Customer = bad | age , employment , salary) ⎟⎠
= β0 + β1WOEage + β2WOEemployment + β3WOE salary
One then commonly applies a scorecard scaling by calculating a
score as a linear function of the log odds, as follows:
Score = offset + factor * log(odds)
Table 8.1

Example Application Scorecard

Characteristic Name

Attribute

Points

Age 1
Age 2

Up to 26
26−35

100
120

Age 3

35−37

185

Age 4

37+
Employed
Unemployed
Up to 500

225

Employment status 1
Employment status 2
Salary 1
Salary 2

501−1,000

90
180
120
140

Salary 3

1,001−1,500

160

Salary 4

1,501−2,000

200

Salary 5

2,001+

240

EXAMPLE APPLICATIONS ◂

163

Assume that we want a score of 600 for odds of 50:1, and a score
of 620 for odds of 100:1. This gives the following:
600 = offset + factor * log(50)
620 = offset + factor * log(100)
The offset and factor then become:
factor = 20/ln(2)
offset = 600 − factor * ln(50)
Once these values are known, the score becomes:
⎛ N
⎞
Score = ⎜ ∑(WOEi * βi ) + β0 ⎟ * factor + offset
⎝ i =1
⎠
⎛ N ⎛
β ⎞⎞
Score = ⎜ ∑ ⎜ WOEi * β i + 0 ⎟ ⎟ * factor + offset
⎝
⎝ i =1
N ⎠⎠
⎛ ⎛
β ⎞
offset ⎞
Score = ⎜ ∑ ⎜ WOEi * β i + 0 ⎟ * factor +
⎝ i =1 ⎝
N⎠
N ⎟⎠
N

Hence, the points for each attribute are calculated by multiplying the
weight of evidence of the attribute with the regression coefficient of the
characteristic, then adding a fraction of the regression intercept, multiplying the result by the factor, and finally adding a fraction of the offset.
In addition to application scorecards, behavioral scorecards are
also typically constructed. These are analytical models that are used
to score the default behavior of an existing portfolio of customers. On
top of the application characteristics, behavioral characteristics, such
as trends in account balance or bureau score, delinquency history,
credit limit increase/decrease, and address changes, can also be used.
Because behavioral scorecards have more data available than application scorecards, their performance (e.g., measured using AUC) will be
higher. Next to debt provisioning, behavioral scorecards can also be
used for marketing (e.g., up/down/cross‐selling) and/or proactive debt
collection. Figure 8.2 gives an example of how a data set for behavioral
scoring is typically constructed.
Both application and behavioral scorecards are then used to calculate the probability of default (PD) for a portfolio of customers. This

164

▸ AN A LY T I C S I N A BI G DATA WO RL D

Checking account

3000
2500
2000
1500
1000
500
0

6
Month

Bureau score

800

Snapshot
t12
Observation
Point

750
700

t24
Good/Bad?

650
600
550
500

6
Month

Number of products purchased
Number of times changed home address
Delinquency history (all credits)
…

Figure 8.2

Constructing a Data Set for Behavioral Scoring

is done by first segmenting the scores into risk ratings and then calculating a historically observed default rate for each rating, which is
then used to project the probability of default (PD) for (typically) the
upcoming year. Figure 8.3 gives an example of how credit risk models
are commonly applied in many bank settings.3

Figure 8.3 Three Level Credit Risk Model

EXAMPLE APPLICATIO NS ◂

165

Other measures that need to be calculated in credit risk modeling are the loss given default (LGD) and exposure at default (EAD).
LGD measures the economic loss expressed as a percentage of the
outstanding loan amount and is typically estimated using linear regression or regression trees. EAD represents the outstanding balance for
on‐balance sheet items (e.g., mortgages, installment loans). For off‐
balance sheet items (e.g., credit cards, credit lines), the EAD is typically
calculated as follows: EAD = DRAWN + CCF * (LIMIT−DRAWN),
whereby DRAWN represents the already drawn balance, LIMIT the
credit limit, and CCF the credit conversion factor, which is expressed
as a percentage between 0 and 1. CCF is typically modeled using either
averages, linear regression, or regression trees.
Once the PD, LGD, and EAD have been estimated, they will be
input into a capital requirements formula provided in the Basel II/III
accord, calculating the necessary amount of capital needed to protect
against unexpected losses.

FRAUD DETECTION
Fraud detection comes in many flavors. Typical examples for which
fraud detection is relevant are: credit card fraud, insurance claim
fraud, money laundering, tax evasion, product warranty fraud, and
click fraud. A first important challenge in fraud detection concerns
the labeling of the transactions as fraudulent or not. A high suspicion does not mean absolute certainty, although this is often used to
do the labeling. Alternatively, if available, one may also rely on court
judgments to make the decision.
Supervised, unsupervised, and social network learning can be used
for fraud detection. In supervised learning, a labeled data set with fraud
transactions is available. A common problem here is the skewness of
the data set because typically only a few transactions will be fraudulent.
Hence, a decision tree already starts from a very pure root node (say,
99 percent nonfraudulent/1 percent fraudulent) and one may not be
able to find any meaningful splits to further reduce the impurity. Similarly, other analytical techniques may have a tendency to simply predict the majority class by labeling each transaction as nonfraudulent.
Common schemes to deal with this are over‐ and undersampling. In

166

▸ A N A LY T I C S I N A BI G DATA WO RL D

oversampling, the fraudulent transactions in the training data set (not
the test data set!) are replicated to increase their importance. In undersampling, nonfraudulent transactions are removed from the training
data set (not test data set!) to increase the weight and importance of the
fraudulent transactions. Both procedures are useful to help the analytical technique in finding a discriminating pattern between fraudulent
and nonfraudulent transactions. Note that it is important to remember
that the test set remains untouched during this. However, if an analytical technique is built using under‐ or oversampling, the predictions it
produces on the test data set may be biased and need to be adjusted.
One way to adjust the predictions is as follows:4
p(C i )
pt (C i ⎢x)
pt (C i )
p(C i ⎢x) =
m p(C )
∑ j =1 pt (Cjj )pt (C j ⎢x)
whereby C i represents the target class (e.g., C1 is fraudulent and C2 is
nonfraudulent), pt (C i | x) represents the probability estimated on the
over‐ or undersampled training data set, pt (C i ) is the prior probability of
class Ci on the over‐ or undersampled training data set, and p(C i ) represents the original priors (e.g., 99/1 percent). The denominator is introduced to make sure that the probabilities sum to one for all classes.
Unsupervised learning can also be used to detect clusters of outlying transactions. The idea here is to build, for example, a SOM and
look for cells containing only a few observations that might potentially
indicate anomalies requiring further inspection and attention.
Finally, social network analysis might also be handy for fraud
detection. Although fraud may be hard to detect based on the available variables, it is often very useful to analyze the relationships
between fraudsters. Rather than a standalone phenomenon, fraud is
often a carefully organized crime. Exploiting relational information
provides some interesting insights in criminal patterns and activities.
Figure 8.4 illustrates a fraud network. Note that this network is constructed around node 1 (in the center of the figure). Nodes in the network that are green are legitimate nodes. Red nodes are fraudulent.
The network visualization gives a good impression of the difference

EXAMPLE APPLICATIO NS ◂

30
28

31
21

19
47

29
10

5
14

22
44

36
34
33

Figure 8.4

23
38

167

42
8

Fraud Network.

Light Gray Nodes Refer to Legitimate Individuals, While Dark Gray Nodes Represent Fraud

in network structure between legitimate and fraudulent nodes. While
legitimate nodes only sparsely connect to each other, fraudulent nodes
are characterized by a dense structure, with many links between all the
members. Such structures have been investigated by Van Vlasselaer,
Meskens, Van Dromme, and Baesens5 and are called spider constructions in the domain of social security fraud. The name spider constructions is derived from their appearance: The fraudulent constructions
look like a dense web in which all nodes are closely connected to each
other. Based on the egonet concept, discussed earlier, both local and
network variables are constructed to characterize each node. Local
variables define the node of interest using only individual characteristics, independent of its surrounding neighbors. Network variables are
dependent on the network structure, and include:
■

Fraudulent degree. In the network domain, the first‐order
degree refers to the number of immediate contacts a node has.
The n‐degree defines the number of nodes the surveyed node

▸ A N A LY T I C S I N A BI G DATA WO RL D

168

can reach in at most n hops. Instead of calculating the overall
degree, one can make a distinction based on the label of each of
the surrounding nodes. For the fraud domain, this means that
the fraudulent first‐order degree corresponds to counting the
number of direct fraudulent neighbors.
■

Triangles. A triangle in a network is defined as a structure in
which three nodes of the network are connected to each other.
Especially triangles containing at least two fraudulent nodes are
a good indicator of potential suspicious activities of the third
node. Nodes that are involved in many suspicious triangles have
a higher probability to commit fraud themselves.

■

Cliques. A clique is an extension of a triangle. Newman (2010)
defines a clique as the maximal subset of the vertices in an undirected network such that every member of the set is connected
by an edge to every other. While fraudulent triangles appear
regularly in a network, fraudulent k‐cliques (with k > 3) will
appear less often. However, such cliques are extremely precise
indicators of future fraud.

Although network variables as such can be very useful in detecting potential future fraud, these characteristics can also be converted
in aggregated variables characterizing each node (e.g., total number
of triangles/cliques, average degree weight, average triangle/clique
weight). Afterward, these network variables should be enriched by
local variables as discussed before. Using all the available attributes,
standard learning techniques like logistic regression, random forests,
and neural networks are able to estimate future fraud based on both
network‐related information and personal information. Such a combined approach exploits all potential information and returns the relevance, in terms of variable weight, of each characteristic.

NET LIFT RESPONSE MODELING
In response modeling, the focus lies on deepening or recovering
customer relationships, or new customer acquisition by means of
targeted or win‐back campaigns. The campaign can be a mail catalog, email, coupon, or A/B or multivariate testing. The purpose is to

EXAMPLE APPLICATIO NS ◂

169

identify the customers most likely to respond based on the following
information:
■

Demographic variables (e.g., age, gender, marital status)

■

Relationship variables (e.g., length of relationship, number of
products purchased)

■

Social network information

■

RFM variables

RFM has been popularized by Cullinan6 as follows:
■

Recency: Time frame (days, weeks, months) since last purchase

■

Frequency: Number of purchases within a given time frame

■

Monetary: Dollar value of purchases

Each of these constructs can be operationalized in various ways;
for example, one can consider the minimum/maximum/average/most
recent monetary value of purchases. The constructs can be used separately or combined into an RFM score by either independent or dependent sorting. For the former (see Figure 8.5), the customer database
is sorted into independent quintiles based on RFM (e.g., recency
quintile 1 is the 20 percent most ancient buyers). The final RFM score

Recency
score

Frequency score

3
2
1

3
Monetary score
5
4
3
2
1

Figure 8.5

Constructing an RFM Score (Independent Sorting)

2
1

170

▸ A N A LY T I C S I N A BI G DATA WO RL D

Figure 8.6

Recency
score

Frequency
score

Monetary
score

335

334

333

332

331

Constructing an RFM Score (Dependent Sorting)

(e.g., 325) can then be used as a predictor for the response model. For
dependent sorting, the customer database is first sorted into quintiles
based on recency (see Figure 8.6). Each recency quintile is then further
divided into frequency quintiles and then into monetary quintiles.
This again yields an RFM score (e.g., 335) that can be used as a predictor for a response model.
A first approach to response modeling is by splitting the previous campaign population into a test group and a control group (see
Figure 8.7). The test group receives the marketing campaign and a
model is built on a training subset and evaluated on a holdout subset.
Traditionally, the impact of such a marketing campaign is measured by
comparing the purchase rate of a test group against the purchase rate
of a control group. If the purchase rate of the test group exceeds the
purchase rate of the control group, the marketing campaign is said to
be effective. Although such methods concentrate on maximizing the
gross purchase rate (i.e., purchase rate test group minus purchase rate
control group), they do not differentiate between different customers
and therefore ignore the net or incremental impact of the campaign.
In general, three types of customers can be distinguished. First, there
are those people who would never buy the product, whether they
are exposed to a marketing offer or not. Targeting these people would
not make any sense because they won’t buy the product anyway. A
second group of customers is those who always buy the product. Targeting these people will cause a profit loss because they will always
buy the product; therefore, offering them a marketing incentive (e.g.,
a discount) will reduce the profit margin. A last category of customers
is the so‐called swing clients. These types of customers will not buy
the product spontaneously, but need to be motivated to take action.

EXAMPLE APPLICATIO NS ◂

171

Previous campaign data

Control

Test

Holdout
data

Training
data

Model
Figure 8.7 Gross Lift Response Modeling

Because they are still undecided on whether to buy the product, a
marketing campaign is especially effective for these people. Focusing
on only these customers will maximize the true impact of the marketing campaign and is the goal of net lift modeling. Net lift modeling
tries to measure the true impact by the incremental purchases, that is,
purchases that are only attributable to the campaign and that would
not be made otherwise.7 Net lift modeling aims at finding a model
such that the difference between the test group purchase rate and the
control group purchase rate is maximized so as to identify the swing
clients (see Figure 8.8). By implementing this methodology, marketers
Previous campaign data

Control

Test

Training
data

Holdout
data

Model
Figure 8.8 Net Lift Response Modeling

▸ A N A LY T I C S I N A BI G D ATA WO RL D

172

Self-selectors
Test group

Converted
swing clients

Y=1

No purchase

Y=0

Self-selectors

Y=1

Control group Swing clients
No purchase

Y=0

Figure 8.9 Observed Target in Net Lift Modeling

not only optimize the true business objective—maximizing profit—but
also gain a better insight in the different customer segments.
In the test and control group, the target will then be observed as
indicated in Figure 8.9.
One could then build a difference score model, as follows:
■

Build a logistic regression model estimating probability of purchase given marketing message, P(purchase|test).

■

Build a logistic regression model estimating probability of purchase given control, P(purchase|control).

■

Incremental score = P(purchase|test)−P(purchase|control).

To further understand the impact of the predictors, one can then
regress the incremental lift scores on the original data.
Another option could be to build only one logistic regression model
with an additional binary predictor specifying whether an observation
belongs to the control or test group. The model can then also include
all possible interaction terms with this binary variable.

CHURN PREDICTION
Customer churn, also called attrition or defection, is the loss of customers. In saturated markets, there are limited opportunities to attract new
customers, so retaining existing customers is essential to profitability
and stability. It is estimated that attracting a new customer costs five to

EXAMPLE APPLICATIO NS ◂

173

six times more than retaining a customer.8 Established customers are
more profitable due to the lower cost to serve them. In addition, brand
loyalty developed over time makes them less likely to churn. Satisfied
customers also serve as word‐of‐mouth advertisement, referring new
customers to the company.
Research on customer churn can take two perspectives: the overall company level and the individual customer level. Identifying the
determinants of churn, or reasons why customers may churn, can
give insight into company‐level initiatives that may reduce the issues
that lead to higher churn. One such study9 performed a survey of the
Korean mobile telephone market. Service attributes such as call quality and tariff level are negatively correlated with churn in that market. Naturally, if it is possible to improve call quality, fewer customers
would be expected to churn. The results of this and similar studies
certainly indicate that management must focus on the quality of attributes that are most important to customers.10 However, continually
improving in these areas may not always be feasible due to cost or
other limitations.
As a complementary approach, switching the focus to the individual
customer level can yield high returns for a relatively low investment.
It is possible to use churn prediction models to identify individual customers who are likely to churn and attempt to prevent them from
leaving the company. These models assign each customer an expected
probability of churn. Then it is relatively straightforward to offer those
customers with the greatest probability a discount or other promotion to encourage them to extend their contract or keep their account
active. In the following section, several techniques and approaches to
churn prediction will be discussed.

Churn Prediction Models
Many well‐known and less common models have been applied to
churn prediction, including decision trees, logistic regression, support
vector machines, Bayesian networks, survival analysis, self‐organizing
maps, and relational classifiers, among others. Both accuracy and comprehensibility are crucial for the decision‐making process, so careful
consideration should be used when choosing a technique. Accurate

174

▸ A N A LY T I C S I N A BI G DATA WO RL D

predictions are perhaps the most apparent goal, but learning the
reasons, or at least the indicators, for churn is also invaluable to the
company. Understanding why a model makes the predictions it does
serves several purposes. Comprehensibility allows for domain experts
to evaluate the model and ensure that it is intuitively correct. In this
way, it can be verified or confirmed by the business. More comprehensible models also offer insight into the correlation between customer
attributes and propensity to churn,11 allowing management to address
the factors leading to churn in addition to targeting the customers
before they decide to churn. Finally, understandable and intuitive
models may be more easily adopted within a company. If managers
are accustomed to making decisions based on their own experience
and knowledge, they will be more inclined to trust predictions made
by a model that is not only comprehensible but also in line with their
own reasoning.
Logistic regression is a statistical classification model that is often
used for churn prediction, either as a model on its own or as a comparison for other models. The coefficients for this model indicate the
correlation between the customer attributes and the probability of
churn. It is a well understood and accepted model both in research
and practice. It is both easy to interpret and provides good results
when compared with other methods. It has been shown to outperform more complex methods in many cases. Decision trees can also be
used for churn prediction. They also offer interpretability and robustness. Neural networks and support vector machines have also been
applied to churn prediction; however, these methods are seen as black
boxes, offering little insight into how the predictions are made. Survival analysis offers the interpretability of logistic regression in the
form of hazard ratios that can be interpreted similarly to odds ratios in
logistic regression. In addition, the target of interest is time‐to‐event
rather than a binary variable. It is therefore possible to make predictions about how long a customer will remain active before they churn.
Relational classifiers can also be used for churn prediction. Homophily in networks is based on the idea that similar individuals are more
likely to interact, and from that it is expected that individuals that are
connected in a network will behave similarly. In churn prediction, if
customers are linked with churners, they may also be likely to churn.

EXAMPLE APPLICATIONS ◂

175

Social network features can also be used in a traditional classifier like
logistic regression or survival analysis. In order to do this, measures of
connectedness can be extracted from the network and used as input
features for the other model.12

Churn Prediction Process
Regardless of the particular technique, churn prediction modeling follows a standard classification process as illustrated in Figure 8.10. The
first step is to define churn for the particular situation. This may be naturally present in the data: contract termination, service cancellation,
or nonrenewal. In other settings, it will not be so clear: A customer
no longer shops at the store or website, or a customer stops purchasing credits. In these cases, the analyst or researcher must choose a
definition of churn that makes sense in the context. One common
solution is to select an appropriate length of time of inactivity on the
account. In the previous examples, a number of days or months without a purchase might define churn. Of course, a customer may not
buy something within that time frame but still return again at a later
date. Setting too short of a time period may lead to nonchurn customers being targeted as potential churners. Too long of a period may
mean churning customers are not identified in a timely manner. In
most cases, a shorter time period may be preferable, if the cost of the
intervention campaign is much lower than the cost of a lost customer.
After defining churn, the original set of customers should be
labeled according to their true churn status. The data set is split for

Unknown
Data

Retention
Campaign
Predictions 8

Training Set
3
1

Define
Churn

No
Campaign

Model
5
Test Set

Figure 8.10

The Churn Prediction Process

Model
Performance

176

▸ A N A LY T I C S I N A BI G DATA WO RL D

validation and the customer attributes from the training set can be
used to train the selected model. The customer attributes from the
test set are then used to compare the model’s results with the actual
churn label. This allows for an evaluation of the model performance.
The model may also be evaluated by domain experts to gauge whether
the predictive attributes seem in line with business knowledge. If the
performance is acceptable, the attributes of current customers can be
entered into the model to predict their churn class. A group of customers with the highest predicted churn probability can then be contacted
with the retention campaign. Other customers who are less likely to
churn are not contacted with the promotion.

RECOMMENDER SYSTEMS
People are influenced by recommendations in their daily decisions.
Salesmen try to sell us the product we like, restaurants are being evaluated and rated, and so on. Recommender systems can support us in
our online commercial activities by suggesting specific items from a
wide range of options. A considerable number of different techniques
are available to build a recommender system, of which the following
are the most important: collaborative filtering, content‐based filtering,
demographic filtering, knowledge‐based filtering, and hybrid filtering. Case studies presenting all these techniques have greatly multiplied in recent years. A lot of these deal with movies,13 tourism,14 and
restaurants.15
In this section, the five main techniques are introduced and followed by some of their advantages and disadvantages. Some other
issues concerning recommender systems are then briefly discussed.

Collaborative Filtering
Collaborative filtering, also called social filtering, has been the approach
that is associated the most with recommender systems. The main idea
is to recommend items based on the opinions of other users. A distinction can be made between user‐based collaborative filtering and
item‐based collaborative filtering. In case of user‐based collaborative
filtering, items will be recommended to a user based on how similar

EXAMPLE APPLICATIO NS ◂

177

users rated these items. When opting for item‐based collaborative filtering, items will be recommended to a user based on how this user
rated similar items. One way to calculate similarity between users or
items is to use a user‐item matrix that contains information on which
user bought what item. Any similarity measure can then be used to
create a similarity matrix (e.g., Pearson correlation and cosine).
To build a collaborative recommender system, ratings are required.
These ratings form the link between a user and an item.16 A distinction can be made between three types of ratings. A scalar rating can
be a number or an ordinal rating. A binary rating consists of two possibilities, such as good or bad. Finally, unary ratings indicate that a user
has had an interaction with an item, such as a click on an item or a
purchase.17 We can distinguish between two types of methods for the
collection of ratings. Explicit ratings can be obtained by requesting a
user to rate a certain item. Implicit ratings are obtained by associating
a rating with a certain action, such as buying an item.18
Typically, neighborhood‐based algorithms are applied, in which the
following three steps can be distinguished.19 First, a similarity measure
is used to calculate similarity between users (in case of a user‐based
algorithm) or items (in case of an item‐based algorithm). Second, a
subset of users or items is selected that functions as the neighborhood
of the active user or item. Third, the algorithm predicts a rating based
on the active user’s or item’s neighborhood, typically giving the highest weight to the most similar neighbors.
As is often the case with analytics, different techniques can be
used to solve the same problem, with their respective advantages and
disadvantages. Three main advantages of collaborative recommender
systems are identified. First, collaborative filtering does not restrict the
type of items to be recommended. It is indeed enough to construct a
matrix linking items to users to start the recommendation. A second
advantage, linked to the first, is that it manages to deliver recommendations to a user even when it is difficult to find out which specific
feature of the item makes it interesting to the user or when there is no
easy way to extract such a feature automatically. A third advantage has
to do with novelty or serendipity: Collaborative filtering is believed to
recommend more unexpected items (that are equally valuable) than
content‐based techniques.20 Although collaborative filtering methods

178

▸ A N A LY T I C S I N A BI G D ATA WO RL D

are the most commonly used techniques because of their power, some
disadvantages or weak points should be noted. First, sparse data can
be a problem for such a technique. A critical mass of ratings is indeed
necessary in order to build meaningful similarity matrices. In cases in
which the items are not frequently bought by the users (e.g., recommending mobile phones or apartments), it may indeed be difficult to
obtain representative neighborhoods, hence lowering the power of the
technique. A second disadvantage is known as the cold start problem,
which means that new items cannot easily be recommended because
they have not been rated yet; therefore, new users cannot easily receive
recommendations because they have not yet rated items. Some minor
disadvantages are, for example, the fact that items purchased a long
time ago may have a substantial impact if few items have been rated,
which may lead to wrong conclusions in a changing environment.
Privacy could also be a problem because collaborative filtering needs
data on users to give recommendations or could generate trust issues
because a user cannot question the recommendation.

Content‐Based Filtering
Content‐based recommender systems recommend items based on two
information sources: features of products and ratings given by users.
Different kinds of data can be encountered, requiring different strategies to obtain usable input. In the case of structured data, each item
consists of the same attributes and the possible values for these attributes are known. It is then straightforward to apply content‐based
approaches. When only unstructured data are available, such as text,
different techniques have to be used in order to learn the user profiles.
Because no standard attributes and values are available, typical problems arise, such as synonyms and polysemous words. Free text can
then be translated into more structured data by using a selection of
free text terms as attributes. Techniques like TF‐IDF (term frequency/
inverse document frequency) can then be used to assign weights to
the different terms of an item. Sometimes, data is semistructured, consisting of some attributes with restricted values and some free text.
One approach to deal with this kind of data is to convert the text into
structured data.21

EXAMPLE APPLICATIO NS ◂

179

When items can be represented in a usable way, machine learning
techniques are applied to learn a user profile. Typically, a classification
algorithm is invoked for each user based on his or her ratings on items
and their attributes. This allows the recommender system to predict
whether a user will like an item with a specific representation. As with
collaborative filtering methods, explicit or implicit ratings are required.
When explicit ratings are considered, the ratings are directly used for
the classification task, whereas implicit ratings can be obtained using
the item–user interactions.
The classification problem mentioned above can be implemented
using a large number of different machine learning techniques. Some
examples are logistic regression, neural networks, decision trees, association rules, and Bayesian networks. Nearest neighbor methods can
also be used to determine the labeled items that are most similar to
a new unlabeled item in order to label this new item based on the
labels of the nearest neighbors. Concerning the similarity metric used
in nearest neighbor methods, Euclidean distance is often used when
data are structured, whereas cosine similarity may prove its use when
the vector space model is applied. Other approaches are linear classifiers, support vector machines, and Naïve Bayes.22
A first advantage of content‐based recommender systems is that
there is no cold start problem for new items. Indeed, new items
(which have not received ratings before) can be recommended,
which was not the case in a collaborative filtering approach. Second, items can also be recommended to users that have unique
preferences. A third important advantage is the possibility to give
an explanation to the user about his or her recommendations, for
example, by means of displaying a list of features that led to the
item being recommended. A fourth advantage is that only ratings
of the active user are used in order to build the profile, which is not
the case for collaborative recommender systems.23 Concerning the
disadvantages, a first limitation is that content‐based techniques are
only suitable if the right data are available. It is indeed necessary to
have enough information about the items to determine whether a
user would like an item or not. The cold start problem for new users
forms a second limitation as well, as old ratings potentially influence
the recommendation too much. Finally, over‐specialization can be a

180

▸ A N A LY T I C S I N A BI G DATA WO RL D

problem because such techniques will focus on items similar to the
previously bought items.

Demographic Filtering
Demographic filtering recommends items based on demographic information of the user. The main challenge is to obtain the data. This can
be explicitly done by asking for information from users such as age,
gender, address, and so on. If this approach is not possible, analytical
techniques could be used to extract information linked to the interactions of the users with the system. A user profile can then be built and
used to recommend items.24
The main advantage of demographic recommender systems is that
there is not always a need for a history of user ratings of the type that is
required in collaborative and content‐based approaches. Segments can
be used in combination with user–item interactions in order to obtain
a high‐level recommender system. Some disadvantages are the cold
start problem for new users and new items, as well as the difficulty in
capturing the data, which is highly dependent on the participation of
the users.

Knowledge‐Based Filtering
Compared with collaborative filtering and content‐based recommender
systems, it is more difficult to briefly summarize the characteristics of
knowledge‐based recommender systems. The main difference with
regard to the other techniques resides in the data sources used. With
this approach, additional inputs consisting of constraints or requirements are provided to the recommender system typically by allowing
a dialog between the user and the system. Knowledge‐based recommender systems can be divided in two main categories: constraint‐
based recommenders and case‐based recommenders. Constraint‐based
recommenders are systems meeting a set of constraints imposed by both
users and the item domain. A model of the customer requirements,
the product properties, and other constraints that limit the possible
requirements is first constructed and formalized. Any technique can
then be used and will have to meet the requirements, or at least

EXAMPLE APPLICATIO NS ◂

181

minimize the violations. When dealing with case‐based recommenders, the goal is to find the item that is most similar to the ones the
user requires. Similarity is then often based on knowledge of the item
domain. The system will then start with an example provided by the
user and will generate a user profile based on it. Based on this user
profile gathering information and additional knowledge sources, recommendations can then be proposed.25
A first advantage of knowledge‐based recommender systems is
that they can be used when there is only limited information about
the user, hence avoiding the cold start problem. Another advantage
is that expert knowledge is used in the recommender system. It is also
possible to function in an environment with complex, infrequently
bought items. In addition, a constraint‐based recommender system
can help customers actively, for example, by explaining products or
suggesting changes in case no recommendation is possible. Concerning
disadvantages, a knowledge‐based recommender system may require
some effort concerning knowledge acquisition, knowledge engineering, and development of the user interface. A second disadvantage is
that it can be difficult when the user is asked to provide the system
with an example if the number of items in the recommendation system is very high. Similarly, it may be difficult or impossible for the user
to provide an example that fits the user’s needs.

Hybrid Filtering
Hybrid recommender systems combine the advantages of content‐
based, knowledge‐based, demographic, and collaborative filtering
recommender systems. The main reason that hybrid recommender
systems have been developed is to avoid the cold start problem.
Burke26 explains seven types of hybrid techniques. A first type is
weighted. In this case, the recommendation scores of several recommenders are combined by applying specific weights. Switching is a second hybrid technique in which recommendations are taken from one
recommender at a time, but not always the same one. A third type of
hybrid technique is mixed. When such a hybrid technique is applied,
recommendations for multiple recommenders are shown to the user.
Feature combination is a fourth type of hybrid technique. In this case,

182

▸ A N A LY T I C S I N A BI G DATA WO RL D

different knowledge sources are used to obtain features, and these are
then given to the recommendation algorithm. A fifth type is feature
augmentation: A first recommender computes the features while the
next recommender computes the remainder of the recommendation.
For example, Melville, Mooney, and Nagarajan27 use a content‐based
model to generate ratings for items that are unrated and then collaborative filtering uses these to make the recommendation. Cascade
is the sixth type of hybrid technique. In this case, each recommender
is assigned a certain priority and if high priority recommenders produce a different score, the lower priority recommenders are decisive.
Finally, a meta‐level hybrid recommender system consists of a first
recommender that gives a model as output that is used as input by
the next recommender. For example, Pazzani28 discusses a restaurant
recommender that first uses a content‐based technique to build user
profiles. Afterward, collaborative filtering is used to compare each
user and identify neighbors. Burke29 states that a meta‐level hybrid is
different from a feature augmentation hybrid because the meta‐level
hybrid does not use any original profile data; the original knowledge
source is replaced in its entirety.

Evaluation of Recommender Systems
Two categories of evaluation metrics are generally considered:30 the
goodness or badness of the output presented by a recommender
system and its time and space requirements. Recommender systems
generating predictions (numerical values corresponding to users’ ratings for items) should be evaluated separately from recommender
systems that propose a list of N items that a user is expected to find
interesting (top‐N recommendation). The first category of evaluation
metrics that we consider is the goodness or badness of the output presented by a recommender system. Concerning recommender systems
that make predictions, prediction accuracy can be measured using
statistical accuracy metrics (of which mean absolute deviation [MAD]
is the most popular one) and using decision support accuracy metrics (of which area under the receiver operating characteristic curve
is the most popular one). Coverage denotes for which percentage of
the items the recommender system can make a prediction. Coverage

EXAMPLE APPLICATIO NS ◂

183

might decrease in case of data sparsity in the user–item matrix. Concerning top‐N
N recommendation, important metrics are recall precision–related measures. Data is first divided in a training set and a
test set. The algorithm runs on the training set, giving a list of recommended items. The concept of “hit set”31 is considered, containing only
the recommended (top‐N) items that are also in the test set. Recall and
precision are then determined as follows:
Recall =

size of hit set
size of test set

Precision =

size of hit set
N

A problem with recall and precision is that usually recall increases
as N is increased, while precision decreases as N is increased. Therefore, the F1 metric combines both measures:32
F1 =

2 * recall * precision
recall + precision

F1 for each user and then taking the average gives the
score of the top‐N recommendation list.
The other category of evaluation metrics is dealing with the
performance of a recommender system in terms of time and space
requirements. Response time is the time that is needed for a system
to formulate a response to a user’s request. Storage requirements can
be considered in two ways: main memory requirement (online space
needed by the system) and secondary storage requirement (offline
space needed by the system).
Additional metrics can also be considered and will depend on
the type of recommender system faced and the domain in which it
is used. For example, it is a common practice in a direct marketing
context to build a cumulative lift curve or calculate the AUC. One
also has to decide whether online or offline evaluations will be made.
Although offline evaluation is typically applied, it is often misleading
because the context of the recommendation is not considered. However, the costs linked with online evaluations are typically higher and
are accompanied by different risks (e.g., bad recommendations may
impact customers’ satisfaction).

184

▸ AN A LY T I C S I N A BI G DATA WO RL D

Examples
Different cases applying recommendation techniques have been
reported, providing the practitioners with best practices and success
stories. Some references are provided in what follows, showing a small
subset of the available cases. A first case that is relevant in the context
of collaborative filtering is Amazon.com. Linden, Smith, and York33
describe the use of recommendation algorithms at Amazon.com.
They see recommendation systems as a type of targeted marketing
because the needs of the customer can be met in a personalized way.
A second case that is relevant in the context of collaborative filtering is PITTCULT, a cultural event recommender based on a network
of trust. In another case, Mooney and Roy34 apply a content‐based
approach on book recommendations. Semistructured text is extracted
from web pages at Amazon.com and text categorization is then applied
to it. Users rate books of the training set, which allows the system
to learn the user profile using a Bayesian learning algorithm. A first
case that is relevant in the context of knowledge‐based recommender
systems is “virtual advisor,” the constraint‐based recommender system proposed by Jannach, Zanker, and Fuchs.35 Virtual advisor is a
knowledge‐based tourism recommender system that has been developed for a premium spa resort in Austria. The authors show that using
a dialog, user requirements and preferences are derived. During the
dialog, the internal user model is analyzed and the next dialog action
is determined. When enough information is gathered about the user’s
requirements and preferences, the system shows the items that meet
the user’s constraints. If necessary, it shows which constraints have
to be relaxed. A second case that is relevant in the context of knowledge‐based recommender systems is Intelligent Travel Recommender
(ITR), discussed by Ricci, Arslan, Mirzadeh, and Venturini.36 ITR is a
case‐based travel advisory system that recommends a travel plan to a
user, starting from some wishes and constraints that this user enters in
the system. The current session is considered a case and it has similarities with cases of other users that are already finished. These previous
cases can have an impact on the recommendation to the users. One
advantage of this approach is that users do not need a login because
the set of past cases that influence the user’s recommendation is based

EXAMPLE APPLICATIO NS ◂

185

on similarity between the user’s case and past cases. A second advantage is that a limited user profile is sufficient, which is not the case
when applying a content‐based approach (as it is then assumed that
users and products share features).

WEB ANALYTICS
The Digital Analytics Association (DAA) defines web analytics as:37
the measurement, collection, analysis, and reporting
of Internet data for the purposes of understanding and
optimizing Web usage.
In what follows, we first elaborate on web data collection and then
illustrate how this can be analyzed.

Web Data Collection
A key challenge in web analytics is to collect data about web visits.38 A
first option here is web server log analysis, which is essentially a server‐
side data collection technique making use of the web server’s logging
functionality. Every HTTP request produces an entry in one or more web
server log files. The log file can then be parsed and processed on a set
schedule to provide useful information. This is illustrated in Figure 8.11.
Common log file formats are:
■

Apache/NCSA log formats: Common Log Format or Combined
Log Format

■

W3C (World Wide Web Consortium) Extended Log File Format
and its Microsoft IIS implementation
Web server(s)
User

HTTP request, for
example, get page
HTML code

Figure 8.11

Web Server Log Analysis

Log file(s)

▸ A N A LY T I C S I N A BI G DATA WO RL D

186

Figure 8.12 Example Log Entry

A log entry (Apache combined log format) typically looks like
Figure 8.12.
The data recorded includes:
■

Remote host: IP address or domain name; helps identify the
geographical location of the client computer

■

Remote log name (“‐”); user name (“‐” if no authentication)

■

Date and time (can include offset from Greenwich Mean Time)

■

HTTP request method (GET or POST)

■

Resource requested

■

Relative to the root directory location on the web server

■

Might include query string (parameters after the ?)“GET/
dutch/shop/detail.html?ProdID=112 HTTP/1.1”

HTTP status code
■

200 range: successful (200 for GET request means requested
resource has been sent)

■

300 range: redirect

■

400 range: client error (404 means not found)

■

500 range: server error

■

Number of bytes transferred

■

Referrer: web page from which user clicked on link to arrive
here

■

“http://www.msn.be/shopping/food/“

■

“http://www.google.com/search?q=buy+wine&hl=
en&lr=“

Browser and platform (user agent)
■

Can also be a search bot, for example, Googlebot

Cookies can also be used for data collection. A cookie is a small
text string that

EXAMPLE APPLICATIO NS ◂

■

187

A web server can send to a visitor’s web browser (as part of its
HTTP response)
The browser can store on the user’s hard disk in the form of a
small text file
The browser sends back unchanged to that server each time a
new request is sent to it (for example, when user visits another
page of the site)

A cookie typically contains a unique user ID along with other customized data, domain, path (specifying from where it can be read),
and expiration date (optional). Cookies can be set and read by (and
their contents shared between) client‐side (e.g., JavaScript) as well as
server‐side (e.g., PHP) scripts. A web server cannot retrieve cookies
from other sites (unless by exploiting vulnerabilities, i.e., cookie stealing). Cookies are typically used for:
■

Implementing virtual shopping carts

■

Remembering user details or providing a customized user
experience without having to log in each time

■

Gathering accurate information about the site’s visitors (session
identification, repeat visitors)

■

Banner ad tracking

A distinction can be made between session and persistent cookies.
A session cookie is used to keep state info for the duration of a visit
and disappears after you close the session/browser. A persistent cookie
is saved to a file and kept long after the end of the session (until the
specified expiration date). Another distinction relates to the originator
of the cookie. A first‐party cookie is set from the same domain that
hosts the web page that is being visited. A third‐party cookie is set by
a web server from another domain, such as an ad network serving
banner ads on the site that is being visited. Third‐party cookies are
typically used to track users across multiple sites and for behavioral
targeting.
Another data collection mechanism in web analytics is page tagging. This is client‐side data collection and usually involves “tagging” a
web page with a code snippet referencing a separate JavaScript file that
deposits and reads a cookie and sends data through to a data collection

188

▸ A N A LY T I C S I N A BI G D ATA WO RL D

Web server(s)
User

request page
HTML code with
JavaScript tag

Data collection
server(s)

When page is loaded, script is run
that collects and sends on data
Figure 8.13

Page Tagging

server. This is illustrated in Figure 8.13. An example Google Analytics
page tag is given in Figure 8.14.
With page tagging, the analytics vendor often provides a hosted
service whereby the client is provided with a web interface to access
reports or run analyses. A popular example of this is Google Analytics.
Tables 8.2 and 8.3 illustrate the advantages and disadvantages, respectively, of page tagging versus web log analysis.
Other techniques have also been suggested for web data collection
but are less commonly used, such as web beacons, packet sniffing, web
server plug‐ins, and/or hybrid solutions.

Web KPIs
Once the data has been collected, it can be analyzed and summarized
into various web key performance indicators (KPIs). Page views are

Figure 8.14

Example Google Analytics Page Tag

EXAMPLE APPLICATIONS ◂

Table 8.2

Advantages of Page Tagging versus Web Server Log Analysis

Page Tagging
■
■
■
■
■
■

■

189

Web Server Log Analysis

Breaks through proxy servers and browser
caching
Tracks client side events (JavaScript, Flash,
etc.)
Easy client‐side collection of outcome data
(custom tags on order confirmation page)
Facilitates real‐time data collection and
processing
Often hosted service available: potential cost
advantages
Data capture separated from web design/
programming: JavaScript code for data
collection can largely be updated by in‐house
analysts or analytics service provider without IT
department having to implement changes
More innovation efforts put in by web analytics
vendors

■

■
■

■
■
■

■

Proxy/caching inaccuracies:
if a page is cached, no record is
logged on your web server
No client‐side event tracking
Most often will choose to integrate
with another database to obtain
additional data
Log files analyzed in batch (unless
server plug‐ins used)
In‐house data collection and
processing
Larger reliance on IT department to
implement changes to capture more
data
Extensive preprocessing required:
“stitch” together log files from
different servers and filter them

the number of times a page (where page is an analyst‐definable unit of
content) was viewed. It is an important building block for other metrics, but it is not that meaningful on its own because we don’t know
whether the customer met his or her purpose after having visited a
page. Also, in today’s web environment, it might not be that straightforward to define a web page unambiguously. The next step is identifying and counting visits or sessions. An example of a visit could be:
index.html ⇒ products.html ⇒ reviews.html ⇒ exit. Sessionization
Table 8.3 Disadvantages of Page Tagging versus Web Server Log Analysis

Page tagging
■

■
■
■

■

Not including correct tags, run‐time errors,
and so on, mean data is lost; cannot go
back
Firewalls and browser privacy/security
settings can hinder data collection
Cannot track search engine bots/spiders/
crawlers (bots do not execute tags)
Less straightforward to capture technical
info such as errors, bandwidth, download
time, and so forth
Loss of control if hosted

Web server log analysis
■
■
■
■
■

Historical data remains available for
reprocessing
Server‐side data collected regardless of
client configuration
Bots/spiders/crawlers show up in log
Designed to automatically capture
technical info
In‐house solution

▸ A N A LY T I C S I N A BI G DATA WO RL D

190

is a procedure for determining which page views are part of the same
visit. In defining sessions, one will make use of a combination of IP
address, user agent, cookies, and/or URI parameters. Once the sessions
have been defined, one could start looking at the visitors. New visitors
are the unique visitors with activity including a first‐ever visit to the
site during a reporting period. Return visitors are the unique visitors
during a reporting period who had also visited the site prior to that
period. This can be interesting to determine loyalty and affinity of visitors. A next obvious question is how long/deep the visits were. This
can be measured with the following metrics:
■

Page views per visit (or also visit depth, page load activity); for
example, the visitor browsed through three different pages

■

Time on page

■

Time on site (also called visit duration or length); for example,
the visit lasted five minutes in total

It is important to note that these metrics should be interpreted in
the appropriate way. For example, a support site might want to solve
the problem quickly and aim for a short time on site and/or call avoidance, whereas a content site might want to get customers engaged and
aim for a longer time on site.
Another very important metric is the bounce rate. It is defined as
the ratio of visits where a visitor left instantly after having seen the
first page. It can be further refined as follows:
■

Bounce rate of the site: ratio of single page view visits (or
bounces) over total visits

■

Bounce rate of a specific page: single page view visits of that
page over number of visits where that page was the entry page

It is also important to consider the referring web page URI because
it also includes search keywords and key phrases for search engine
traffic sources. Other interesting measures are:
■

Most viewed pages (top content, popular pages)

■

Top entry pages

■

Top exit pages (leakage)

■

Top destinations (exit links)

EXAMPLE APPLICATIONS ◂

191

Finally, a very important metric is the conversion rate. A conversion is a visitor performing an action that is specified as a useful outcome considering the purpose of the site. The conversion rate is then
defined as the percentage of visits or of unique visitors for which we
observed the action (e.g., order received, lead collected, newsletter
sign up). It is hereby important to combine the conversion rate also
with other outcome data, such as sales price, revenue, ROI, and so on.
For a checkout process, one could consider the following metrics:
■

Cart abandonment rate = 1 − number of people who start
checkout/total Add to Cart clicks

■

Checkout abandonment rate = 1 − number of people who
complete checkout/number of people who start checkout

It is important to note that small improvements in these metrics
can usually lead to substantial revenue gains.
The average visits or days to purchase is a pan‐session metric giving insight into how long it takes people to buy from your website (or
submit a lead).

Turning Web KPIs into Actionable Insights
Ultimately, it is the purpose to transform the metrics discussed earlier
into actionable insights. Each metric should be compared in time to
see whether there are any significant changes. For example, popular
referrers are disappearing, new referrers come in, top five referrers
changed, top destinations changed, and so forth. Trend analysis is very
useful here. It is important to verify whether there is an upward/downward trend, or any seasonalities or daily/weekly/monthly patterns to
observe. This is illustrated in Figure 8.15 for the conversion rate.
Dashboards will be used to effectively monitor and communicate
the web KPIs. They often provide intuitive indicators such as gauges,
stoplights, and alerts and can be personalized.
KPI
This week
Conversion rate
1.6%
…
Figure 8.15

Last week
2.0%

Monitoring the Conversion Rate

Percent change
–20%

▸ A N A LY T I C S I N A BI G DATA WO RL D

192

Benchmarking can also be very useful to compare internal web
KPIs against industry standards. Popular benchmark service providers
are Fireclick and Google Analytics’s benchmarking service.
Segmentation is also very important in order to turn web KPIs into
actionable insights. Any measure can be broken down into segments
of interest and aggregate (total, proportion, average) numbers can
be computed per segment. For example, one could segment bounce/
conversion rates by:
■

Top five referrers

■

Search traffic or not

■

Geographical region

■

Acquisition strategy (i.e., direct marketing, PPC, SEO/organic
search, email marketing, newsletter, affiliates)

This can be very efficiently supported by means of OLAP facilities
to perform interactive analysis of large volumes of web KPI data from
multiple dimensions.
Site search reports are also very useful because they provide a basic
understanding of the usage of the internal search engine. This is a basic
form of market research because the users tell you exactly what they
are looking for. It is interesting to consider the following:
■

■

Site search usage
■

How much is the search function used?

■

What keywords are used most?

Site search quality
■

Calculate bounce rate for site search (% search exits)

Navigation Analysis
Navigation analysis allows us to understand how users navigate
through the website.
Path analysis gives insight into frequent navigation patterns. It
analyzes, from a given page, which other pages a group of users visit
next in x percent of the times. Note, however, that this assumes that
the users follow a linear path, which is not always the case.

EXAMPLE APPLICATIO NS ◂

193

A funnel plot focuses on a predetermined sequence (e.g., a check
out process) and measures entry/abandonment at each stage.
A page overlay/click density analysis shows clicks or other metrics
(e.g., bounce/conversion rates) overlaid directly on actual pages such
that one can traverse through the website as a group of users typically
navigates through it. Heat maps then have colors indicating the click
frequencies.
Again, it is important to combine all these plots with segmentation
to give actionable insights.

Search Engine Marketing Analytics
Web analytics can also be used to measure the efficiency of search
engine marketing. Two types of search engine marketing are search
engine optimization (SEO) and pay per click (PPC). In SEO, the
purpose is to improve organic search results in a search engine (e.g.,
Google, Yahoo!) without paying for it. This can be accomplished by
carefully designing the website. In PPC, one pays a search engine for
a link/ad to the website to appear in the search results. The link/ad is
then listed depending on the search engine algorithm, the bid, and the
competitor’s bids. Popular examples are Google AdWords and Yahoo!
Search Marketing. SEO efforts can be measured as follows:
■

Inclusion ratio = number of pages indexed/number of pages on
your website. Note that sometimes you do not want pages to be
indexed, to avoid users arriving too deep within a website.

■

Robot/crawl statistics report. See how frequently your website
is being visited by search engine robots and how deep they get.
Note that this should be done based on seb log analysis, since
robots do not run JavaScript page tags.

■

Track inbound links by using www.mysite.com in Google.

■

Google webmaster tools that show, for the most popular search
keywords or phrases that have returned pages from your site,
the number of impressions or user queries for which your
website appeared in the search results and the number of users
who actually clicked and came to your website.

▸ A N A LY T I C S I N A BI G D ATA WO RL D

194
■

Track rankings for your top keywords/key phrases.

■

See whether keywords link to your most important pages.

PPC efforts can be tracked as follows:
■

Reports that differentiate bid terms versus search terms when
users enter site through PPC campaign (e.g., bid term is “laptop”
but search term is “cheap laptops”)

■

Analyze additional data obtained about ad impressions, clicks,
cost

■

Keyword position report (for example, AdWords position report)
■

Specifies position your ad was in when clicked

■

Can show any metric (e.g., unique visitors, conversion rate,
bounce rate) per position

A/B and Multivariate Testing
The purpose here is to set up an experiment whereby different pages
or page elements are shown to randomly sampled visitors. Example
pages that could be considered are landing page (first page of a visit),
page in checkout process, most popular page(s), or pages with high
bounce rates.
In A/B testing, one tests two alternative versions of a web page
on a random sample of visitors and compares against a control group
(who gets the original page). This is illustrated in Figure 8.16.

Random
sample
50%
Clicked
on link
to page

Figure 8.16

A/B Testing

Conversion
rate
Original

2.0%

25%

Version
A

1.9%

25%

Version
B

3.0%

Conversion
page

Test significance!

EXAMPLE APPLICATIO NS ◂

195

X1: headline

X2: sales
copy

X3: image
(e.g., “hero
shot”)

X4: button text
Figure 8.17 Multivariate Testing

Multivariate testing aims at testing more than one element on a
page at the same time (see Figure 8.17). Note that one can also test
price sensitivity or different product bundles, which requires integration with back‐end data sources.
Parametric data analysis can then be used to understand the effect
of individual page elements and their interactions on a target measure
of interest (e.g., bounce or conversion rate). Also, techniques from
experimental design can be used to intelligently decide on the various
page versions to be used.

SOCIAL MEDIA ANALYTICS
With the rising popularity of the web, people are closer connected to
each other than ever before. While it only has been a few years since
people communicated with each other on the street, the demographic
boundaries are fading away through the recently trending online
communication channels. The marginal effect of traditional word‐of‐
mouth advertising is replaced by the enormous spread of information
and influence through the wires of the World Wide Web. Web users
have been putting billions of data online on websites like Facebook
and MySpace (social network sites), Twitter (microblog site), YouTube
and DailyMotion (multimedia‐sharing), Flickr and ShutterFly (photo
sharing), LinkedIn and ZoomInfo (business‐oriented social network
site), Wikipedia and Open Directory Profound (user‐generated encyclopedia), Reddit (content voting site), and many others.
Users are no longer reluctant to share personal information about
themselves, their friends, their colleagues, their idols, and their political

196

▸ A N A LY T I C S I N A BI G DATA WO RL D

preferences with anybody who is interested in them. Nowadays, with
the booming rise of mobile applications, web users are 24/7 connected
to all kinds of social media platforms, giving real‐time information
about their whereabouts. As such, a new challenging research domain
arises: social media analytics. While these data sources offer invaluable
knowledge and insights in customer behavior and enable marketers
to more carefully profile, track, and target their customers, crawling
through such data sources is far from evident because social media
data can take immense magnitudes never seen before.
From a sales‐oriented point of view, social media offers advantages
for both parties in the business–consumer relationship. First, people
share thoughts and opinions on weblogs, microblogs, online forums,
and review websites, creating a strong effect of digital word‐of‐mouth
advertising. Web users can use others’ experience to gain information and make purchase decisions. As such, consumers are no longer falling for transparent business tricks of a sales representative, but
they are well‐informed and make conscious choices like true experts.
Public opinions are volatile. Today’s zeroes are tomorrow’s heroes.
Companies are forced to keep offering high‐quality products and services, and only a small failure can have disastrous consequences for
the future. Keeping one step ahead of the competition is a tough and
intensive process, especially when regional competitors are also able
to enter the game. On a large scale, the main competitors for an industry used to consists of the big players of the market, while local businesses were too small and playing together with the big guys required
capital‐intensive investments. The Internet changed the competitive
environment drastically, and consumers can easily compare product
and service characteristics of both local and global competitors.
Although the merciless power of the public cannot be underestimated, companies should embrace and deploy social media data.
People trust social media platforms with their personal data and interests, making it an invaluable data source for all types of stakeholders.
Marketers who are searching for the most promising and profitable
consumers to target are now able to capture more concrete consumer
characteristics, and hence develop a better understanding of their customers. Zeng39 described social media as an essential component of
the next‐generation business intelligence platform. Politicians and

EXAMPLE APPLICATIO NS ◂

197

governmental institutions can get an impression of the public opinion
through the analysis of social media. During election campaigns, studies claim that political candidates with a higher social media engagement got relatively more votes within most political parties.40 Social
media analytics is a select tool to acquire and propagate one’s reputation. Also, nonprofit organizations such as those in the health sector
benefit from the dissemination power of social media, anticipating, for
example, disease outbreaks, identifying disease carriers, and setting up
a right vaccination policy.41
Social media analytics is a multifaceted domain. Data available on
social media platforms contain diverse information galore, and focusing
on the relevant pieces of data is far from obvious and often unfeasible.
While certain social media platforms allow one to crawl publicly accessible data through their API (application programming interface), most
social networking sites are protective toward data sharing and offer
built‐in advertisement tools to set up personalized marketing campaigns. This is briefly discussed in the first subsection. The next subsections introduce some basic concepts of sentiment and network analysis.

Social Networking Sites: B2B Advertisement Tools
A new business‐to‐business (B2B) billion‐dollar industry is launched
by capturing users’ information in social network websites, enabling
personalized advertising and offering services for budget and impact
management.
Facebook Advertising42 is a far‐evolved marketing tool with an
extensive variety of facilities and services (see Figure 8.18). Depending
on the goal of the advertising campaign, Facebook Advertising calculates the impact and spread of the digital word‐of‐mouth advertising.
Facebook Advertising not only supports simple marketing campaigns
such as increasing the number of clicks to a website (click rate) or page
likes (like rate) and striving for more reactions on messages posted by
the user (comment and share rate), but also more advanced options
like mobile app engagement (download and usage rate) and website
conversion (conversion rate) are provided. The conversion rate of a
marketing campaign refers to the proportion of people who undertake
a predefined action. This action can be an enrollment for a newsletter,

198

▸ A N A LY T I C S I N A BI G DATA WO RL D

Figure 8.18 Determining Advertising Objective in Facebook Advertising

leaving an email address, buying a product, downloading a trial version, and so on, and is specific for each marketing campaign. Facebook
measures conversion rates by including a conversion‐tracking pixel
on the web page where conversion will take place. A pixel is a small
piece of code communicating with the Facebook servers and tracking
which users saw a web page and performed a certain action. As such,
Facebook Advertising matches the users with their Facebook profile
and provides a detailed overview of customer characteristics and the
campaign impact.
Facebook Advertising allows users to create personalized ads and
target a specific public by selecting the appropriate characteristics in
terms of demographics, interests, behavior, and relationships. This is
shown in Figure 8.19. Advertisements are displayed according to a bidding system, where the most eye‐catching spots of a page are the most
expensive ones. When a user opens his or her Facebook page, a virtual
auction decides which ad will be placed where on the page. Depending
on the magnitude and the popularity of (a part of) the chosen audience,
Facebook suggests a bidding amount. A safer solution is to fix a maximum bid amount in advance. The higher the amount of the bid, the
higher the probability of getting a good ad placement. Notice, however,
that the winning bid does not necessarily have to pay the maximum bid
amount. Only when many ads are competing do ad prices rise drastically. As such, the price of an ad differs depending on the target user.

EXAMPLE APPLICATIO NS ◂

Figure 8.19

199

Choosing the Audience for Facebook Advertising Campaign

The business‐oriented social networking site LinkedIn offers similar services as Facebook. The LinkedIn Campaign Manager43 allows
the marketer to create personalized ads and to select the right customers. Compared to Facebook, LinkedIn Campaign Managers offers services to target individuals based on the characteristics of the companies
they are working at and the job function they have (see Figure 8.20).
While Facebook Advertising is particularly suitable for Business‐to‐
Consumer (B2C) marketing, LinkedIn Campaign Manager is aimed at
advertisements for Business‐to Business (B2B) and Human Resource
Management (HRM) purposes.
As most tools are self-explanatory, the reader must be careful when
deploying these advertisement tools since they may be so user friendly
that the user no longer realizes what he/she is actually doing with them.
Make sure that you specify a maximum budget and closely monitor all
activities and advertisement costs, definitely at the start of a marketing campaign. A small error can result in a cost of thousands or even
millions of dollars in only a few seconds. Good knowledge of all the
facilities is essential to pursue a healthy online marketing campaign.

200

▸ A N A LY T I C S I N A BI G DATA WO RL D

Figure 8.20

LinkedIn Campaign Manager

Sentiment Analysis
Certain social media platforms allow external servers to capture data
from a portion of the users. This gateway for external applications is
called the API. An API has multiple functions. It offers an embedded
interface to other programs. For example, the Twitter API44 can be
used on other sites to identify visitors by their Twitter account. Integrated tweet fields and buttons on web pages allow users to directly
post a reaction without leaving the web page. Like buttons are directly
connected to your Facebook page through the Facebook API45 and
immediately share the like with all of your friends. However, APIs
often permit external servers to connect and mine the publicly available data. Undelimited user‐generated content like text, photos, music,
videos, and slideshows is not easy to interpret by computer‐controlled
algorithms.
Sentiment analysis and opinion mining focus on the analysis of
text and determining the global sentiment of the text. Before the
actual sentiment of a text fragment can be analyzed, text should be

EXAMPLE APPLICATIONS ◂

201

Figure 8.21 Sentiment Analysis for Tweet

preprocessed in terms of tag removal, tokenization, stopword removal,
and stemming. Afterward, each word is associated with a sentiment.
The dominant polarity of the text defines the final sentiment.
Because text contains many irrelevant words and symbols, unnecessary tags are removed from the text, such as URLs and punctuation marks. Figure 8.21 represents an example of a tweet. The link in
the tweet does not contain any useful information, thus it should be
removed for sentiment analysis. The tokenization step converts the
text into a stream of words. For the tweet shown in Figure 8.21, this
will result in:
Data Science / rocks / excellent / book / written / by / my / good /
friends / Foster Provost / and / Tom Fawcett / a / must / read
In a next step, stopwords are detected and removed from the
sentence. A stopword
d is a word in a sentence that has no informative
meaning, like articles, conjunctions, prepositions, and so forth. Using
a predefined machine‐readable list, stopwords can easily be identified
and removed. Although such a stoplist can be constructed manually,
words with an IDF (inverse document frequency) value close to zero
are automatically added to the list. These IDF values are computed
based on the total set of text fragments that should be analyzed. The
more a word appears in the total text, the lower its value. This gives:
Data Science / rocks / excellent / book / written / good / friends /
Foster Provost / Tom Fawcett / read
Many variants of a word exist. Stemming converts each word back
to its stem or root: All conjugations are transformed to the corresponding verb, all nouns are converted to their singular form, and adverbs
and adjectives are brought back to their base form. Applied to the previous example, this results in:
Data Science / rock / excellent / book / write / friend / Foster Provost /
Tom Fawcett / read

▸ A N A LY T I C S I N A BI G D ATA WO RL D

202

Each word has a positive (+), negative (−) or neutral (o) polarity.
Again, algorithms use predefined dictionaries to assign a sentiment to
a word. The example contains many positive and neutral words, as
shown below:
Data Science / rock / excellent / book / write / friend / Foster Provost / Tom Fawcett / read
o

The overall sentiment of the above tweet is thus positive. Although
this procedure could easily capture the sentiment of a text fragment,
more advanced analysis techniques merge different opinions from
multiple users together and are able to summarize global product or
service affinity, as well as assign a general feeling toward neutral‐
polarized words.

Network Analytics
Instead of analyzing user‐generated content, network analytics focuses
on the relationships between users on social media platforms. Many
social media platforms allow the user to identify their acquaintances.
Five types of relationships can be distinguished:46
1. Friends. There is a mutual positive relationship between two
users. Both users know each other, and acknowledge the association between them.
2. Admirers. A user receives recognition from another user, but
the relationship is not reciprocal.
3. Idols. A user acknowledges a certain positive connectedness
with another user, but the relationship is not reciprocal.
4. Neutrals. Two users do not know each other and do not communicate with each other.
5. Enemies. There is a negative relationship between two users.
Both users know each other, but there is a negative sphere.
Although in most social networking sites only friendship relationships are exploited, Twitter incorporates admirers (followers) and idols
(followees) by enabling users to define the people they are interested
in. Admirers receive the tweets of their idols. Enemy relationships are
not common in social networking sites, except for EnemyGraph.47 The

EXAMPLE APPLICATIO NS ◂

203

power of social network sites depends on the true representation of
real‐world relationships between people. Link prediction is one subdomain of network analytics where one tries to predict which neutral
links are actually friendship, admirer, or idol relationships. Tie strength
prediction is used to determine the intensity of a relationship between
two users.
Homophily, a concept from sociology, states that people tend to
connect to other similar people and they are unlikely to connect with
dissimilar people. Similarity can be expressed in terms of the same
demographics, behavior, interests, brand affinity, and so on. As such,
in networks characterized by homophily, people connected to each
other are more likely to like the same product or service. Gathering the
true friendship, admirer, and idol relationships between people enables
marketers to make more informed decisions for customer acquisition
and retention. An individual surrounded by many loyal customers
has a high probability of being a future customer. Customer acquisition projects should identify those high‐potential customers based
on the users’ neighborhoods and focus their marketing resources on
them. This is shown in Figure 8.22(a). However, a customer whose
friends have churned to the competition is likely to be a churner as
well, and should be offered additional incentives to prevent him or her

(a)

(b)

Figure 8.22 Social Media Analytics for Customer Acquisition (a) and Retention (b). Grey
nodes are in favor of a specific brand, black nodes are brand‐averse.

204

▸ A N A LY T I C S I N A BI G DATA WO RL D

from leaving. Similar to customer acquisition, these customers can be
detected using relational information available on social media platforms. This is shown in Figure 8.22(b). Influence propagates through
the network. The aforementioned analysis techniques focus on the
properties of the direct neighborhood (one hop). Although direct associates contain important information, more advanced algorithms focus
on influence propagation of the whole network, revealing interesting
patterns impossible to detect with the bare eye.
Although social media analytics nowadays is indispensable in
companies’ market research projects, it is highly advised to verify the
regional, national, and international privacy regulations before starting (see privacy section). In the past, some companies did not comply
with the prevailing privacy legislation and risked very steep fines.

BUSINESS PROCESS ANALYTICS
In recent years, the concept of business process management (BPM)
has been gaining traction in modern companies.48 Broadly put, the
management field aims to provide an encompassing approach in order
to align an organization’s business processes with the concerns of every
involved stakeholder. A business process is then a collection of structured, interrelated activities or tasks that are to be executed to reach a
particular goal (produce a product or deliver a service). Involved parties in business processes include, among others, managers (“process
owners”), who expect work to be delegated swiftly and in an optimal
manner; employees, who desire clear and understandable guidelines
and tasks that are in line with their skillset; and clients who, naturally, expect efficiency and quality results from their suppliers. Figure 8.23 gives an example business process model for an insurance
claim intake process shown in the business process modeling language
(BPMN) standard. Numerous visualization forms exist to design and
model business processes, from easy flowchart‐like diagrams to complex formal models.
Put this way, BPM is oftentimes described as a “process optimization” methodology and is therefore mentioned together with related
quality control terms such as total quality management (TQM), six
sigma efforts, or continuous process improvement methodologies.

205

Figure 8.23

Claim
intake
Evaluate
claim

Example Business Process Model

Review
policy
Propose
settlement

Reject claim

Approve damage
payment

Calculate new
premium
Close
claim

206

▸ A N A LY T I C S I N A BI G DATA WO RL D

Figure 8.24

Business Process Management Lifecycle

However, this description is somewhat lacking. Indeed, one significant focal point of BPM is the actual improvement and optimization of
processes, but the concept also encompasses best practices toward the
design and modeling of business processes, monitoring (consider for
instance compliance requirements), and gaining insights by unleashing analytical tools on recorded business activities. All these activities
are grouped within the “business process lifecycle,” starting with the
design and analysis of a business process (modeling and validation), its
configuration (implementation and testing), its enactment (execution
and monitoring), and finally, the evaluation, which in turn leads again
to the design of new processes (see Figure 8.24).

Process Intelligence
It is mainly in the last part of the BPM life cycle (i.e., evaluation)
where the concepts of process analytics and process intelligence fit in.
Just as with business intelligence (BI) in general, process intelligence
is a very broad term describing a plethora of tools and techniques, and
can include anything that provides information to support decision
making.
As such, just as with traditional (“flat”) data‐oriented tools, many
vendors and consultants have defined process intelligence to be synonymous with process‐aware query and reporting tools, oftentimes
combined with simple visualizations in order to present aggregated
overviews of a business’s actions. In many cases, a particular system

EXAMPLE APPLICATIO NS ◂

207

will present itself as being a helpful tool toward process monitoring
and improvement by providing KPI dashboards and scorecards, thus
presenting a “health report” for a particular business process. Many
process‐aware information support systems also provide online analytical processing (OLAP) tools to view multidimensional data from
different angles and to drill down into detailed information. Another
term that has become commonplace in a process intelligence context
is business activity monitoring (BAM), which refers to real‐time monitoring of business processes and immediate reaction if a process displays a
particular pattern. Corporate performance managementt (CPM) is another
popular term for measuring the performance of a process or the organization as a whole.
Although all the tools previously described, together with all the
three‐letter acronym jargon, are a fine way to measure and query many
aspects of a business’s activities, most tools unfortunately suffer from
the problem that they are unable to provide real insights or uncover
meaningful, newly emerging patterns. Just as for non‐process‐related
data sets (although reporting, querying, aggregating and drilling, and
inspecting dashboard indicators are perfectly reasonable for operational day‐to‐day management), these tools all have little to do with
real process analytics. The main issues lies in the fact that such tools
inherently assume that users and analysts already know what to look
for. That is, writing queries to derive indicators assumes that one
already knows the indicators of interest. As such, patterns that can
only be detected by applying real analytical approaches remain hidden. Moreover, whenever a report or indicator does signal a problem,
users often face the issue of then having to go on a scavenger hunt
in order to pinpoint the real root cause behind the problem, working
all the way down starting from a high‐level aggregation toward the
source data. Figure 8.25 provides an example of a process intelligence
dashboard.
Clearly, a strong need is emerging to go further than straightforward
reporting in today’s business processes and to start a thorough analysis
directly from the avalanche of data that is being logged, recorded, and
stored and is readily available in modern information support systems,
leading us to the areas of process mining and analytics.

208

▸ A N A LY T I C S I N A BI G D ATA WO RL D

Figure 8.25 Example Process Intelligence Dashboard
Source: http://dashboardmd.net.

Process Mining and Analytics
In the past decade, a new research field has emerged, denoted as “process
mining,” which positions itself between BPM and traditional data mining. The discipline aims to provide a comprehensive set of tools to provide process‐centered insights and to drive process improvement efforts.
Contrary to business intelligence approaches, the field emphasizes a
bottom‐up approach, starting from real‐life data to drive analytical tasks.
As previously stated, process mining builds on existing approaches,
such as data mining and model‐driven approaches, but is more than
just the sum of these components. For example, as seen previously,
traditional existing data mining techniques are too data‐centric to provide a solid understanding of the end‐to‐end processes in an organization, whereas business intelligence tools focus on simple dashboards
and reporting. It is exactly this gap that is narrowed by process mining
tools, thus enabling true business process analytics.
The most common task in the area of process mining is called process discovery, in which analysts aim to derive an as‐is process model
starting from the data as it is recorded in process‐aware information
support systems instead of starting from a to‐be descriptive model, and

EXAMPLE APPLICATIO NS ◂

209

trying to align the actual data to this model. A significant advantage of
process discovery is the fact that only a limited amount of initial data
is required to perform a first exploratory analysis.
Consider, for example, the insurance claim handling process as it
was previously depicted. To perform a process discovery task, we start
our analysis from a so‐called “event log”: a data table listing the activities that have been executed during a certain time period, together
with the case (the process instance) to which they belong. A simple
event fragment log for the insurance claim handling process might look
as depicted in Table 8.4. Activities are sorted based on the starting time.
Note that multiple process instances can be active at the same moment
in time. Note also that the execution of some activities can overlap.
Based on real‐life data as it was stored in log repositories, it is possible to derive an as‐is process model that provides an overview of
how the process was actually executed. To do this, activities are sorted
based on their starting time. Next, an algorithm iterates over all process cases and creates “flows of work” between the activities. Activities
that follow each other distinctly (no overlapping start and end times)
Table 8.4 Example Insurance Claim Handling Event Log

Case
Identifier

Start Time

Completion Time

Activity

Z1001

8‐13‐2013 09:43:33

8‐13‐2013 10:11:21

Claim intake

Z1004

8‐13‐2013 11:55:12

8‐13‐2013 15:43:41

Claim intake

Z1001

8‐13‐2013 14:31:05

8‐16‐2013 10:55:13

Evaluate claim

Z1004

8‐13‐2013 16:11:14

8‐16‐2013 10:51:24

Review policy

Z1001

8‐17‐2013 11:08:51

8‐17‐2013 17:11:53

Propose settlement

Z1001

8‐18‐2013 14:23:31

8‐21‐2013 09:13:41

Calculate new premium

Z1004

8‐19‐2013 09:05:01

8‐21‐2013 14:42:11

Propose settlement

Z1001

8‐19‐2013 12:13:25

8‐22‐2013 11:18:26

Approve damage payment

Z1004

8‐21‐2013 11:15:43

8‐25‐2013 13:30:08

Approve damage payment

Z1001

8‐24‐2013 10:06:08

8‐24‐2013 12:12:18

Close claim

Z1004

8‐24‐2013 12:15:12

8‐25‐2013 10:36:42

Calculate new premium

Z1011

8‐25‐2013 17:12:02

8‐26‐2013 14:43:32

Claim intake

Z1004

8‐28‐2013 12:43:41

8‐28‐2013 13:13:11

Close claim

Z1011

8‐26‐2013 15:11:05

8‐26‐2013 15:26:55

Reject claim

210

▸ A N A LY T I C S I N A BI G DATA WO RL D

will be put in a sequence. When the same activity is followed by different activities over various process instances, a split is created. When
two or more activities’ executions overlap in time, they are executed
in parallel and are thus both flowing from a common predecessor.
After executing the process discovery algorithm, a process map
such as the one depicted in Figure 8.26 can be obtained (using the

3
Claim intake
3
1

1
Review Policy
1

Evaluate claim
1
1

1
Propose settlement
2
2
Calculate new premium
2

Reject claim
1
1

Approve damage payment
2
2

2
Close claim
2
2

Figure 8.26 Example of a Discovered Process Map Annotated with Frequency Counts

EXAMPLE APPLICATIONS ◂

211

Disco software package). The process map can be annotated with various information, such as frequency counts of an activity’s execution.
Figure 8.27 shows the same process map now annotated with performance‐based information (mean execution time). Note that, together
with solid filtering capabilities, visualizations such as these provide an
excellent means to perform an exploratory analytics task to determine

Claim intake
8.6 hrs
4.3 hrs

Calculate new premium
44.6 hrs

Reject claim
15.8 mins

70.2 hrs

Propose settlement
29.8 hrs
45.4 hrs

27.6 mins

Review Policy
66.7 hrs

Evaluate claim
68.4 hrs
24.2 hrs

27.6 mins

5d
43 hrs

Approve damage payment
35 d
3.1 d

59 hrs
Close claim
77.8 mins

Figure 8.27 Example Process Map Annotated with Performance Information

212

▸ A N A LY T I C S I N A BI G DATA WO RL D

bottlenecks and process deviations, compared to having to work with
flat data–based tools (e.g., analyzing the original event log table using
spreadsheet software).
As can be seen from the figures, process discovery provides an
excellent means to perform an initial exploratory analysis of the data
at hand, showing actual and true information. This allows practitioners to quickly determine bottlenecks, deviations, and exceptions in
the day‐to‐day workflows.
Other, more advanced process discovery tools exist to extract
other forms of process models. We discuss here the so‐called Alpha
algorithm, which was put forward by Wil van der Aalst as one of the
first formal methods to extract process models containing split/join
semantics, meaning that this discovery algorithm aims to discover
explicitly which tasks occur in parallel; in the process maps shown in
Figures 8.26 and 8.27, only high level “flows” between activities are
depicted, which provides a solid, high‐level overview of the process
but can be made more specific.49
The Alpha algorithm assumes three sets of activities: Tw is the
set containing all activities, Ti is the set containing all activities
that occur as a starting activity in a process instance (e.g., “claim
intake”), and To is the set of all activities that occur as an ending
activity in a process instance (e.g., “reject claim” and “close claim”).
Next, basic ordering relations are determined, starting with . It
is said that a b holds when activity a directly precedes b in some
process instance. Based on this set of orderings, it is said that a → b
(sequence) holds if and only if a b ∧ b / a . Also, a # b (exclusion) if and only if a / b ∧ b / a and a || b (inclusion) if and only if
a b ∧ b a . Based on this set of relations, a “footprint” of the log
can be constructed, denoting the relation between each pair of
activities, as depicted in Figure 8.28.

a b
a # →
b ← #
c ← ||

c
→
||
#

Figure 8.28 Footprint Construction in the Alpha Algorithm

EXAMPLE APPLICATIONS ◂

213

Based on this footprint, it is possible to derive semantic relations
between activities:
■

a → b : a and b follow in sequence

■

a → b ∧ a → c ∧ b # c : choice between b or c after a

■

a → c ∧ b → c ∧ a # b : c can follow both after a or b

■

a → b ∧ a → c ∧ b || c : b and c are executed both in parallel after a

■

a → c ∧ b → c ∧ a || b : c follows after both a and b are executed in
parallel

The resulting process model is then shown as a “workflow net,” a
specific class of Petri nets (see Figure 8.29). Note that the parallelism
between “calculate new premium” and “approve damage payment”
and the choice between “review policy” and “evaluate claim” are now
depicted in an explicit manner.
Process discovery is not the only task that is encompassed by process mining. One other particular analytical task is denoted as conformance checking, and this aims to compare an event log as it was executed
in real life with a given process model (which could be either discovered or given). This then allows one to quickly pinpoint deviations and
compliance problems.
Consider once more our example event log. When “replaying”
this event log on the original BPMN model, we immediately see some
deviations occurring. Figure 8.30 depicts the result after replaying
process instance Z1004. As can be seen, the required activity “evaluate claim” was not executed in this trace, causing a compliance
problem for the execution of “propose settlement.” Conformance
checking thus provides a powerful means to immediately uncover
root causes behind deviations and compliance violations in business
processes.

Review
Policy
Claim
Intake

Calculate
New Premium
Propose
Settlement

Evaluate
Claim
Reject
Claim

Figure 8.29

Workflow Net for the Insurance Case

Approve
Damage Payment

Close
Claim

214

Figure 8.30

Review Policy
Conformant

Conformance Checking

Checking instance Claim Intake
Z1004
Conformant

Claim
Intake

Review
Policy

Propose
Settlement

Reject Claim

Calculate
New Premium

Approve
Pay Damages
Close
Claim

Evaluate Claim Propose Settlement Approve Pay Damages Calculate New Premium Close Claim
Non-executed
Execution Violation
Conformant
Conformant
Conformant
Expected
Dubious
Dubious
Dubious

Evaluate
claim

Missing
Activity

Execution
Violation

EXAMPLE APPLICATIO NS ◂

215

This concludes our overview of process mining and its common
analytics tasks. Note that there exist various other process analytics
tasks as well. The following list enumerates a few examples:
■

Rule‐based property verification of compliance checking (e.g.,
in an audit context: verifying whether the four‐eyes principle
was applied when needed)

■

Taking into account additional data, other than case identifiers,
activity names, and times; for instance, by also incorporating
information about the workers having executed the tasks

■

Combining process mining with social analytics; for instance,
to derive social networks explaining how people work together

■

Combining process discovery with simulation techniques to
rapidly iterate on what‐if experiments and to predict the impact
of applying a change in the process

Although Process Mining mainly entails descriptive tasks, such
as exploring and extracting patterns, techniques also exist to support
decision makers in predictive analytics. One particular area of interest has been the prediction of remaining process instance durations
by learning patterns from historical data. Other approaches combine
process mining with more traditional data mining techniques, which
will be described further in the next section.

Coming Full Circle: Integrating with Data Analytics
The main difference between process analytics (process mining) and
data analytics lies in the notion that process mining works on two
levels of aggregation. At the bottom level, we find the various events
relating to certain activities and other additional attributes. By sorting
these events and grouping them based on a case identifier, as done by
process discovery, it becomes possible to take a process‐centric view on
the data set at hand. Therefore, many process mining techniques have
been mainly focusing on this process‐centric view, while spending less
time and effort to aim to produce event‐granular information.
Because of this aspect, it is strongly advisable for practitioners to
adopt an integrated approach by combining process‐centric techniques
with other data analytics, as was discussed throughout this book. We

216

▸ A N A LY T I C S I N A BI G DATA WO RL D

Figure 8.31 Example Spaghetti Model

provide a practical example describing how to do so by integrating process mining and analytics with clustering and predictive decision trees.
To sketch out the problem context, consider a process manager
trying to apply process discovery to explore a very complex and flexible business process. Workers are given many degrees of freedom to
execute particular tasks, with very few imposed rules on how activities
should be ordered. Such processes contain a high amount of variability,
which leads process discovery techniques to extract so‐called spaghetti
models (see Figure 8.31).
Clearly, this is an undesirable scenario. Although it is possible to filter
out infrequent paths or activities, one might nevertheless prefer to get a
good overview on how people execute their assigned work without hiding low‐frequency behavior that may signify both problematic, rare cases
and also possible strategies to optimize the handling of certain tasks that
have not become commonplace yet. This is an important note to keep in
mind for any analytics task: Extracting high‐frequency patterns is crucial
to get a good overview and derive main findings, but even more important is to analyze data sets based on the impact of patterns—meaning
the low frequent patterns can nevertheless uncover crucial knowledge.
Clustering techniques exist to untangle spaghetti models, such
as the process model shown, into multiple smaller models, which all
capture a set of behavior and are more understandable. One such technique, named ActiTraC, incorporates an active learning technique to
perform the clustering, meaning that clusters are created by iteratively
applying a process discovery algorithm on a growing number of process instances until it is determined that the derived process model
becomes too complex and a new cluster is instantiated.50 Figure 8.32
shows how the event log previously shown can be decomposed into
the following sublogs with associated discovered process models. The

EXAMPLE APPLICATIO NS ◂

217

A spaghetti model is obtained after applying process discovery on a flexible, unstructured process:

Log is clustered in smaller sublogs based on common behavior:
Cluster 1 capturing 74 percent
of process instances
Cluster 2 capturing 11 percent
of process instances
Cluster 3 capturing 4 percent
of process instances

Unclustered log model

Cluster 4 capturing 11 percent of remaining,
nonfitting, low-frequent process instances
2

Cluster characteristics are analysed to build predictive decision tree:
Cluster 1

Mean completion time: 3.3 days
Mean number of workers involved: 2
Involved product types: P201, P202
...

Cluster 2

Mean completion time: 4.5 days
Mean number of workers involved: 5
Involved product types: P203
...

Cluster 3

Mean completion time: 32.4 days
Mean number of workers involved: 12
Involved product types: P204
...

Cluster 4

Attribute 1

Attribute 2

Cluster 1

Cluster 2

Attribute 3

Cluster 3

Cluster 4

Mean completion time: 11.7 days
Mean number of workers involved: 7
Involved product types: P205, P206, P207
...

Characteristics of new instances can be predicted:
Predicted cluster: 2

New Process Instance

Involved product type: P203

...

Figure 8.32

Clustering of Process Instances

Expected completion time: 4.5 days
Expected amount of involved workers: 5

218

▸ A N A LY T I C S I N A BI G DATA WO RL D

discovered process models show an easier‐to‐understand view on
the different types of behavior contained in the data. The last cluster
shown here contains all process instances that could not be captured in
one of the simpler clusters and can thus be considered a “rest” category
containing all low‐frequency, rare process variants (extracted with
ActiTraC plugin in ProM software package).
After creating a set of clusters, it is possible to analyze these further
and to derive correlations between the cluster in which an instance
was placed and its characteristics. For example, it is worthwhile to
examine the process instances contained in the final “rest” cluster to
see whether these instances exhibit significantly different run times
(either longer or shorter) than the frequent instances.
Since it is now possible to label each process instance based on the
clustering, we can also apply predictive analytics in order to construct a
predictive classification model for new, future process instances, based
on the attributes of the process when it is created. Figure 8.33 shows
how a decision tree can be extracted for an IT incident handling process. Depending on the incident type, involved product, and involved
department, it is possible to predict the cluster with which a particular
instance will match most closely and, as such, derive expected running time, activity path followed, and other predictive information.

Incident type

“Bug report”

Involved
product

“Product A,” “Product E,” “Product F”

Cluster 1

“Product B,” “Product C,” “Product D”

Cluster 4

“Feature request”

Department

“Finance,” “HR,” “Sales”

“Marketing,” “Management”

“Other”

Figure 8.33

Example Decision Tree for Describing Clusters

Cluster 2

Cluster 3

Cluster 4

Standard
behavior,
average
runtime of
one day
“Deviating”
cluster, long
running
time,
varying
activity
sequence
Standard
behavior,
average
runtime of
three days

Standard
behavior,
average
runtime of
two days

EXAMPLE APPLICATIO NS ◂

219

Decision makers can then apply this information to organize an efficient division of workload.
By combining predictive analytics with process analytics, it is
now possible to come full circle when performing analytical tasks in
a business process context. Note that the scope of applications is not
limited to the example previously described. Similar techniques have
also been applied, for example, to:
■

Extract the criteria that determine how a process model will
branch in a choice point

■

Combine process instance clustering with text mining

■

Suggest the optimal route for a process to follow during its execution

■

Recommend optimal workers to execute a certain task51 (see
Figure 8.34)

As a closing note, we draw attention to the fact that this integrated
approach does not only allow practitioners and analysts to “close the

Figure 8.34

Example Decision Tree for Recommending Optimal Workers

Source: A. Kim, J. Obregon, and J. Y. Jung, “Constructing Decision Trees from Process Logs for
Performer Recommendation,” First International Workshop on Decision Mining & Modeling for
Business Processes (DeMiMoP’13), Beijing, China, August 26–30, 2013.

220

▸ A N A LY T I C S I N A BI G DATA WO RL D

loop” regarding the set of techniques being applied (business analytics, process mining, and predictive analytics), but also enables them
to actively integrate continuous analytics within the actual process
execution. This is contrary to being limited to a post‐hoc exploratory
investigation based on historical, logged data. As such, process
improvement truly becomes an ongoing effort, allowing process owners to implement improvements in a rapid and timely fashion, instead
of relying on reporting–analysis–redesign cycles.

NOTES
1. T. Van Gestel and B. Baesens, Credit Risk Management: Basic Concepts: Financial Risk
Components, Rating Analysis, Models, Economic and Regulatory Capitall (Oxford University
Press, 2009); L. C. Thomas, D. Edelman, and J. N. Crook, Credit Scoring and Its
Applications (Society for Industrial and Applied Mathematics, 2002).
2. B. Baesens et al., “Benchmarking State of the Art Classification Algorithms for Credit
Scoring,” Journal of the Operational Research Society 54, no. 6 (2003): 627–635.
3. T. Van Gestel and B. Baesens, Credit Risk Management: Basic Concepts: Financial Risk
Components, Rating Analysis, Models, Economic and Regulatory Capitall (Oxford University
Press, 2009).
4. M. Saerens, P. Latinne, and C. Decaestecker, “Adjusting the Outputs of a Classifier to
New a Priori Probabilities: A Simple Procedure,” Neural Computation 14, no. 1 (2002):
21–41.
5. V. Van Vlasselaer et al., “Using Social Network Knowledge for Detecting Spider Constructions in Social Security Fraud,” in Proceedings of the 2013 IEEE/ACM International
Conference on Advances in Social Network Analysis and Mining (Niagara Falls, 2013).
IEEE Computer Society.
6. G. J. Cullinan, “Picking Them by Their Batting Averages’ Recency—Frequency—
Monetary Method of Controlling Circulation,” Manual Release 2103 (New York:
Direct Mail/Marketing Association, 1977).
7. V. S. Y. Lo, “The True Lift Model—A Novel Data Mining Approach to Response
Modeling in Database Marketing,” ACM SIGKDD Explorations Newsletterr 4, no. 2
(2002).
8. W. Verbeke et al., “Building Comprehensible Customer Churn Prediction Models with Advanced Rule Induction Techniques,” Expert Systems with Applications 38
(2011): 2354–2364.
9. H.‐S. Kim and C.‐H. Yoon, “Determinants of Subscriber Churn and Customer Loyalty
in the Korean Mobile Telephony Market,” Telecommunications Policy 28 (2004):
751–765.
10. S. Y. Lam et al., “Customer Value, Satisfaction, Loyalty, and Switching Costs: An
Illustration from a Business‐to‐Business Service Context, Journal of the Academy
of Marketing Science 32, no. 3 (2009): 293–311; B. Huang, M. T. Kechadim, and
B. Buckley, “Customer Churn Prediction in Telecommunications,” Expert Systems
with Applications 39 (2012): 1414–1425; A. Aksoy et al., “A Cross‐National Investigation of the Satisfaction and Loyalty Linkage for Mobile Telecommunications Services
across Eight Countries,” Journal of Interactive Marketing 27 (2013): 74–82.

EXAMPLE APPLICATIO NS ◂

221

11. W. Verbeke et al., “Building Comprehensible Customer Churn Prediction Models with Advanced Rule Induction Techniques,” Expert Systems with Applications 38
(2011): 2354–2364.
12. Q. Lu and L. Getoor, “Link‐Based Classification Using Labeled and Unlabeled Data,”
in Proceedings of the ICML Workshop on The Continuum from Labeled to Unlabeled Data
(Washington, DC: ICML, 2003).
13. C. Basu, H. Hirsh, and W. Cohen, “Recommendation as Classification: Using Social
and Content‐based Information in Recommendation,” in Proceedings of the Fifteenth
National/Tenth Conference on Artificial Intelligence/Innovative Applications of Artificial
Intelligence, American Association for Artificial Intelligence (American Association for Artificial Intelligence, Menlo Park, CA, 1998), 714–720; B. N. Miller et al., “Movielens
Unplugged: Experiences with an Occasionally Connected Recommender System,”
in Proceedings of the 8th International Conference on Intelligent User Interfaces (New York,
2003), 263–266. ACM New York, NY, USA.
14. D. Jannach, M. Zanker, and M. Fuchs, “Constraint‐Based Recommendation in
Tourism: A Multi‐Perspective Case Study,” Journal of IT & Tourism 11, no. 2 (2009):
139–155; F. Ricci et al., “ITR: A Case‐based Travel Advisory System,” in Proceeding
of the 6th European Conference on Case Based Reasoning, ECCBR 2002 (Springer‐Verlag
London, UK 2002), 613–627.
15. M. J. Pazzani, “A Framework for Collaborative, Content‐Based and Demographic
Filtering,” Artificial Intelligence Review
w 13, no. 5–6 (1999): 393–408.
16. J. Schafer et al., Collaborative Filtering Recommender Systems, The Adaptive Web (2007),
291–324. Springer‐Verlag Berlin, Heidelberg 2007.
17. Ibid.
18. Ibid.
19. F. Cacheda et al., “Comparison of Collaborative Filtering Algorithms: Limitations of
Current Techniques and Proposals for Scalable, High‐Performance Recommender
System,” ACM Transactions on the Web 5, no. 1 (2011): 1–33.
20. J. Schafer et al., Collaborative Filtering Recommender Systems, The Adaptive Web (2007),
291–324. Springer‐Verlag Berlin, Heidelberg 2007.
21. M. Pazzani and D. Billsus, Content‐Based Recommendation Systems, The Adaptive Web
(2007), 325–341. Springer‐Verlag Berlin, Heidelberg 2007.
22. Ibid.
23. R. J. Mooney and L. Roy, “Content‐Based Book Recommending Using Learning
for Text Categorization,” in Proceedings of the Fifth ACM Conference on Digital Libraries (2000), 195–204; M. De Gemmis et al., “Preference Learning in Recommender
Systems,” in Proceedings of Preference Learning (PL‐09), ECML/PKDD‐09 Workshop
(2009). ACM, New York, NY, USA 2000.
24. M. Pazzani and D. Billsus, Content‐Based Recommendation Systems, The Adaptive Web
(2007), 325–341. Springer‐Verlag Berlin, Heidelberg 2007.
25. A. Felfernig and R. Burke, “Constraint‐Based Recommender Systems: Technologies
and Research Issues,” in Proceedings of the 10th International Conference on Electronic
Commerce, ICEC ’088 (New York: ACM, 2008), 1–10.
26. R. Burke, “Hybrid Web Recommender Systems” in The Adaptive Web (Springer
Berlin/Heidelberg, 2007), 377–408. Springer Berlin Heidelberg.
27. P. Melville, R. J. Mooney, and R. Nagarajan, “Content‐Boosted Collaborative
Filtering for Improved Recommendations,” in Proceedings of the National Conference on
Artificial Intelligence (2002), 187–192. American Association for Artificial Intelligence
Menlo Park, CA, USA 2002.

222

▸ A N A LY T I C S I N A BI G D ATA WO RL D

28. M. Pazzani and D. Billsus, Content‐Based Recommendation Systems, The Adaptive Web
(2007), 325–341.
29. R. Burke, “Hybrid Web Recommender Systems” in The Adaptive Web (Springer
Berlin/Heidelberg, 2007), 377–408. Springer Berlin Heidelberg.
30. E. Vozalis and K. G. Margaritis, “Analysis of Recommender Systems’ Algorithms,” in
Proceedings of The 6th Hellenic European Conference on Computer Mathematics & Its Applications (HERCMA) (Athens, Greece, 2003). LEA Publishers Printed in Hellas, 2003.
31. Ibid.
32. Ibid.
33. G. Linden, B. Smith, and J. York, “Amazon.com Recommendations: Item‐to‐item
Collaborative Filtering,” Internet Computing, IEEE
E 7, no. 1 (2003): 76–80.
34. R. J. Mooney and L. Roy, “Content‐Based Book Recommending Using Learning
for Text Categorization,” in Proceedings of the Fifth ACM Conference on Digital Libraries
(2000), 195–204.
35. D. Jannach, M. Zanker, and M. Fuchs, “Constraint‐Based Recommendation in
Tourism: A Multi‐Perspective Case Study,” Journal of IT & Tourism 11, no. 2 (2009):
139–155.
36. Ricci et al., “ITR: A Case‐based Travel Advisory System,” in Proceeding of the 6th
European Conference on Case Based Reasoning, ECCBR 2002 (Springer‐Verlag London,
UK 2002), 613–627.
37. www.digitalanalyticsassociation.org
38. A. Kaushik, Web Analytics 2.0 (Wiley, 2010).
39. D. Zeng et al., “Social Media Analytics and Intelligence,” Intelligent Systems, IEEE
E 25,
no. 6 (2010): 13–16.
40. R. Effing, J. Van Hillegersberg, and T. Huibers, Social Media and Political Participation: Are Facebook, Twitter and YouTube Democratizing Our Political Systems? Electronic
Participation (Springer Berlin Heidelberg, 2011): 25–35.
41. A. Sadilek, H. A. Kautz, and V. Silenzio, “Predicting Disease Transmission from Geo‐
Tagged Micro‐Blog Data,” AAAI 2012.
42. www.facebook.com/advertising
43. www.linkedin.com/advertising
44. http://dev.twitter.com
45. http://developers.facebook.com
46. P. Doreian and F. Stokman, eds., Evolution of Social Networks (Routledge, 1997).
47. http://enemygraph.com
48. W. M. P. Van Der Aalst, Process Mining: Discovery, Conformance and Enhancement of
Business Processes (Springer Verlag, 2011).
49. W. M. P. Van Der Aalst, A. J. M. M. Weijters, and L. Maruster, “Workflow Mining:
Discovering Process Models from Event Logs,” IEEE Transactions on Knowledge and
Data Engineering 16, no. 9 (2004): 1128–1142; W. M. P. Van Der Aalst, Process Mining:
Discovery, Conformance and Enhancement of Business Processes (Springer Verlag, 2011).
50. J. De Weerdt et al., “Active Trace Clustering for Improved Process Discovery,” IEEE
Transactions on Knowledge and Data Engineering 25, no. 12 (2013): 2708–2720.
51. A. Kim, J. Obregon, and Y. Jung, “Constructing Decision Trees from Process Logs for
Performer Recommendation,” in Proceedings of the DeMiMop’13 Workshop, BPM 2013
Conference (Bejing, China, 2013). Springer.

About the Author

Bart Baesens is an associate professor at KU Leuven (Belgium) and
a lecturer at the University of Southampton (United Kingdom). He
has done extensive research on analytics, customer relationship management, web analytics, fraud detection, and credit risk management
(see www.dataminingapps.com). His findings have been published in
well‐known international journals (e.g., Machine Learning, Management
Science, IEEE Transactions on Neural Networks, IEEE Transactions on Knowledge and Data Engineering, IEEE Transactions on Evolutionary Computation,
and Journal of Machine Learning Research) and presented at top international conferences. He is also co‐author of the book Credit Risk Management: Basic Concepts (Oxford University Press, 2008). He regularly
tutors, advises, and provides consulting support to international firms
with respect to their analytics and credit risk management strategy.

223

INDEX

A
A priori property, 94
A/B testing, 168, 194–195
Accessibility, 151
Accountability principle, 157
Accuracy ratio (AR), 77, 139
Accuracy, 150, 151, 173
Action plan, 144
ActiTrac, 216
Activation function, 49
Active learning, 216
Actuarial method, 110
Adaboost, 65–66
Alpha algorithm, 212
Alter, 129
Amazon, 184
Analytical model requirements, 9–10
Analytics, 7–9
process model, 4–6
Anatomization, 158
ANOVA, 30, 47
Apache/NCSA, 185
API, 200
Apriori algorithm, 90, 93
Area under the ROC curve (AUC), 75,
117, 139, 182
benchmarks, 76
Assignment decision, 42
Association rules, 87–93
extensions, 92–93
mining, 90–91
multilevel, 93
post processing, 92
Attrition, 172
B
Backpropagation learning, 50
B2B advertisement tools, 197
Backtesting, 134–146
classification models, 136–142
clustering models, 143–144

framework, 144–146
policy, 144
regression models, 143
Bagging, 65
Bar chart, 18
Basel II, 36, 161
Basel III, 36, 161
Basic nomenclature, 4
Behavioral scoring, 2
Behavioral targeting, 187
Believability, 151
Benchmark
expert–based, 147
external, 146
Benchmarking, 146–149, 192
Best matching unit (BMU), 100
Betweenness, 121
Bias term, 48
Bid term, 194
Bigraph, 130–132
Binary rating, 177
Binning, 24
Binomial test, 140
Black box, 55
techniques, 52
Board of Directors, 159
Boosting, 65
Bootstrapping procedures, 73
Bounce rate, 190
Box plot, 21
Brier score, 139
Bureau-based inference, 16
Business activity monitoring (BAM), 207
Business expert, 6
Business intelligence, 206
Business process analytics, 204–220
Business process lifecycle, 206
Business process management (BPM),
204
Business process modeling language
(BPMN), 204

225

226

▸ INDEX

Business process, 204
Business relevance, 9, 133
Business-to-Business (B2B), 199
Business-to-Consumer (B2C), 199
C
C4.5 (See5), 42
Capping, 23
Cart abandonment rate, 191
CART, 42
Case-based recommenders, 180
Categorization, 24–28
Censoring, 105
interval, 106
left, 105
right, 105
Centrality measures, 121
CHAID, 42
Champion-challenger, 147
Checkout abandonment rate, 191
Chief Analytics Officer (CAO), 159
Chi-squared, 43
analysis, 25
Churn prediction, 134, 172–176
models, 173
process, 175
Churn
active, 35
expected, 36
forced, 36
passive, 36
Classification accuracy, 74
Classification error, 74
Classing, 24
Click density, 193
Clique, 168
Cloglog, 42
Closeness, 121
Clustering, 216
Clustering, Using and Interpreting,
102–104
Coarse classification, 24
Cold start problem, 177, 179, 180, 181
Collaborative filtering, 176–178
Collection limitation principle, 156
Collective inference, 123–124, 128
Column completeness, 150
Combined log format, 185
Commercial software, 153
Common log format, 185
Community mining, 122

Competing risks, 116
Completeness, 150, 151
Compliance, 213
Component plane, 101
Comprehensibility, 133, 173, 174
Conditional density, 108
Confidence, 87, 89, 94–95
Conformance checking, 213
Confusion matrix, 74
Conjugate gradient, 50
Consistency, 152
Constraint-based recommenders, 180
Content based filtering, 178–180
Continuous process improvement, 204
Control group, 170
Conversion rate, 191, 197
Convex optimization, 64
Cookie stealing, 187
Cookies, 186
first-party, 187
persistent, 187
session, 187
third-party, 187
Corporate governance, 159
Corporate performance management
(CPM), 207
Correlational behavior, 123
Corruption perception index (CPI), 101
Coverage, 182
Cramer’s V, 31
Crawl statistics report, 193
Credit conversion factor (CCF), 165
Credit rating agencies, 146
Credit risk modeling, 133, 146, 161–
165
Credit scoring, 15, 36, 58
Cross-validation, 72
Leave-one-out, 72
Stratified, 72
Cumulative accuracy profile (CAP),
77, 137
Customer acquisition, 203
Customer attrition, 35
Customer lifetime value (CLV), 4,
35–36
Customer retention, 203
Cutoff, 74
D
Dashboard, 191, 207
Data cleaning, 5

INDEX ◂

Data mining, 7
Data poolers, 14
Data publisher, 157
Data quality, 149–152
dimensions, 150
principle, 156
Data science, 7
Data set split up, 71
Data sparsity, 183
Data stability, 136, 143
Data warehouse administrator, 6
Database, 6
Decimal scaling, 24
Decision trees, 42–48, 65, 67, 104, 218
multiclass, 69
Decompositional techniques, 52
Defection, 172
Degree, 121
Demographic filtering, 180
Dendrogram, 98–99, 123
Department of Homeland Security, 156
Dependent sorting, 169
Development sample, 71
Deviation index, 136
Difference score model, 172
Digital analytics association (DAA), 185
Digital dashboard, 144
Disco, 211
Distance measures
Euclidean, 97, 100
Kolmogorov-Smirnov, 79, 137
Mahalanobis, 80
Manhattan, 97
Distribution
Bernoulli, 39
Binomial, 140
Exponential, 111–112
Generalized gamma, 113
Normal, 140
Weibull, 112
Divergence metric, 80
Document management system, 159
Documentation test, 159
Doubling amount, 41
E
Economic cost, 10, 133
Edge, 119
Effects
external, 135
internal, 135

227

Ego, 129
Egonet, 129, 167
Ensemble
methods, 64–65
model, 66
Entropy, 43
Epochs, 50
Equal frequency binning, 25
Equal interval binning, 25
Estimation sample, 71
Evaluating predictive models, 71–83
Event log, 209
Event time distribution, 106
cumulative, 107
discrete, 107
Expert-based data, 14
Explicit rating, 177
Exploratory analysis, 5
Exploratory statistical analysis, 17–19
Exposure at default (EAD), 165
Extended log file format, 185
F
F1 metric, 183
Facebook advertising, 197
Fair Information Practice Principles
(FIPPs), 156
Farness, 121
Feature space, 61, 62, 64
Featurization, 126
FICO score, 14, 146
Fidelity, 55
Filters, 29
Fireclick, 192
Fisher score, 30
Four-eyes principle, 215
Fraud detection, 3, 36, 133, 165–168
Fraudulent degree, 167
Frequent item set, 89, 90
F-test, 144
Funnel plot, 193
G
Gain, 45
Garbage in, garbage out (GIGO), 13, 149
Gartner, 1
Generalization, 158
Geodesic, 121
Gini coefficient, 77
Gini, 43
Girvan-Newman algorithm, 123

228

▸ INDEX

Global minimum, 50
Goodman-Kruskal ϒ, 147
Google AdWords, 193
Google Analytics benchmarking
service, 192
Google analytics, 188
Google webmaster tools, 193
Googlebot, 186
Graph theoretic center, 121
Graph
bipartite, 131
unipartite, 130
Gross response, 36
Gross purchase rate, 170
Grouping, 24
Guilt by association, 124
H
Hazard function, 107
cumulative, 113
Hazard ratio, 115–116
Hazard shapes
constant, 108
convex bathtub, 108
decreasing, 108
increasing, 108
Hidden layer, 49
Heat map, 193
Hidden neurons, 51
Hierarchical clustering, 96–99
agglomerative, 96
divisive, 96
Histogram, 18, 21, 143
Hit set, 183
Hold out sample, 71
Homophily, 124, 129, 174, 203
Hosmer-Lemeshow test, 141
HTTP request, 185
HTTP status code, 186
Hybrid filtering, 181–182
I
Implicit rating, 177
Impurity, 43
Imputation, 19
Inclusion ratio, 193
Incremental impact, 170
Independent sorting, 169
Individual participation principle, 157
Information value, 30, 136
Input layer, 49

Insurance claim handling process, 209
Insurance fraud detection, 4
Intelligent Travel Recommender (ITR),
184
Interestingness measure, 92
Interpretability, 9, 52, 55, 64, 117, 133,
151
Interquartile range, 22
Intertransaction patterns, 94
Intratransaction patterns, 94
IP address, 186
Item-based collaborative filtering, 176
Iterative algorithm, 50
Iterative classification, 128
J
Job profiles, 6–7
Justifiability, 9, 133
K
Kaplan Meier analysis, 109–110
KDnuggets, 1, 2, 153
Kendall’s τ, 147
Kernel function, 61–62
Keyword position report, 194
Kite network, 121–122
K-means clustering, 99
Knowledge diamonds, 5
Knowledge discovery, 7
Knowledge-based filtering, 180–181
L
Lagrangian multipliers, 62
Lagrangian optimization, 60–61, 64
Landing page, 194
Leaf nodes, 42
Legal experts, 6
Levenberg-Marquardt, 50
Life table method, 110
Lift curve, 76
Lift measure, 87, 91–92
Likelihood ratio statistic, 110
Likelihood ratio test, 110, 113–114
Linear decision boundary, 41
Linear kernel, 62
Linear programming, 58
Linear regression, 38
Link characteristic
binary-link, 126
count-link, 126
mode-link, 126

INDEX ◂

Linkage
average, 98
centroid, 98
complete, 98
single, 97
Ward’s, 98
Local minima, 50
Link prediction, 203
LinkedIn campaign manager, 199
Local model, 123
Log entry, 186
Log file, 185
Log format, 185
Logistic regression, 39, 48, 126, 161
cumulative, 68
multiclass, 67–69
relational, 126
Logit, 40, 41
Log-rank test, 110
Loopy belief propagation, 128
Lorenz curve, 77
Loss given default (LGD), 35, 37, 165
M
Mantel-Haenzel test, 110
Margin, 6, 58
Market basket analysis, 93
Markov property, 124
Matlab, 153
Maximum likelihood, 41, 68–69, 112
nonparametric, 109
Mean absolute deviation (MAD), 143,
182
Mean squared error (MSE), 46, 83,
143
Medical diagnosis, 133
Memoryless property, 111
Microsoft Excel, 155
Microsoft, 153
Min/max standardization, 24
Missing values, 19–20
Model
board, 159
calibration, 143
monitoring, 134
performance, 55
ranking, 136, 143
Monotonic relationship, 147
Model design and documentation,
158–159
Moody’s RiskCalc, 42

229

Multiclass
classification techniques, 67
confusion matrix, 80
neural networks, 69–70
support vector machines, 70
Multilayer perceptron (MLP), 49
Multivariate outliers, 20
Multivariate testing, 168, 194–195
Multiway splits, 46
N
Navigation analysis, 192–193
Neighbor-based algorithm, 177
Neighborhood function, 101
Net lift response modeling, 168–172
Net response, 36
Network analytics, 202–204
Network model, 124
Neural network, 48–57, 62
Neuron, 48
Newton Raphson optimization, 113
Next best offer, 3, 93
Node, 119
Nonlinear transformation function, 49
Nonmonotonicity, 25
Notch difference graph, 80
O
Objectivity, 151
Odds ratio, 41
OLAP, 18, 192
OLTP, 14
One-versus-all, 70
One-versus-one, 70
Online analytical processing (OLAP), 207
Open source, 153
Openness principle, 157
Operational efficiency, 10, 133
Opinion mining, 200
Organization for Economic
Cooperation and Development
(OECD), 156
Outlier detection and treatment, 20–24
Output layer, 49
Overfitting, 45, 66
Oversampling, 166
Ownership, 159
P
Packet sniffing, 188
Page overlay, 193

230

▸ INDEX

Page tagging, 187
Page view, 188
Pairs
concordant, 148
discordant, 148
Partial likelihood estimation, 116
Partial profile, 155
Path analysis, 192
Pay per click (PPC), 193
Pearson correlation, 29, 83, 143
Pedagogical rule extraction, 55
Pedagogical techniques, 52
Performance measures for classification
models, 74–82
Performance measures for regression
models, 83
Performance metrics, 71
Permutation, 158
Perturbation, 158
Petri net, 213
Pie chart, 17
Pittcult, 184
Pivot tables, 27
Polynomial kernel, 62
Polysemous word, 178
Population completeness, 150
Posterior class probabilities, 136
Power curve, 77
Precision, 183
Predictive and descriptive analytics, 8
Principal component analysis, 67
Privacy Act, 156
Privacy preserving data mining, 157
Privacy, 7, 15, 155–158, 178, 204
Probabilistic relational neighbor
classifier, 125–126
Probability of default (PD), 163, 164
Probit, 42
Process discovery, 208
Process intelligence, 206–208
Process map, 210
Process mining, 208–215
Product limit estimator, 109
Proportional hazards
assumption, 116
hazards regression, 114–116
Publicly available data, 15
Purpose specification principle, 156
Q
Quadratic programming problem,
60–61

Qualitative checks, 144
Quasi-identifier, 157
R
R, 153
Radial basis function, 62
Random forests, 65–67
Recall, 183
Receiver operating characteristic
(ROC), 75, 117, 137
Recommender systems, 93, 176–185
Recursive partitioning algorithms
(RPAs), 42
Referrer, 186
Regression tree, 46, 65
Regulation, 10, 156
Regulatory compliance, 32, 133
Reject inference, 16
Relational neighbor classifier, 124
Relaxation labeling, 128
Relevancy, 151
Reputation, 151
Response modeling, 2, 36, 133,
168
Response time, 183
Retention modeling, 133
RFM (recency, frequency, monetary),
17, 169
Risk rating, 164
Robot report, 193
Robot, 193
Roll rate analysis, 37
Rotation forests, 67
R-squared, 83, 143
Rule
antecedent, 89
consequent, 89
extraction, 52
set, 46
S
Safety safeguards principle, 157
Sample variation, 134
Sampling, 15–16
bias, 15
Gibbs, 128
stratified, 16
Scatter plot, 18, 83, 143
SAS, 153
Scalar rating, 177
Schema completeness, 150
Scorecard scaling, 162

INDEX ◂

Scorecard, 161, 207
Application, 161
Behavioral, 163
Scoring, 136
Scree plot, 98–99
Search Engine Marketing Analytics,
193–194
Search engine optimization (SEO), 193
Search term, 194
Security, 151
Segmentation, 32–33, 48, 95–96, 192
Self-organizing map (SOM), 100–102
Senior management, 159
Sensitivity, 74
analysis, 92
Sequence rules, 94–95
Sentiment analysis, 200–202
Session, 187, 189
Sessionization, 189
Sigmoid transformation, 23
Sign operator, 60
Similarity measure, 177
Site search, 192
quality, 192
report, 192
usage, 192
Six sigma, 204
Small data sets, 72
Social filtering, 176
Social media analytics, 3, 195–204
Social network, 215
learning, 123–124, 165
metrics, 121–123
Sociogram, 120
Software, 153–155
commercial, 153
open-source, 153
Sparseness property, 62
Spaghetti model, 216
Sparse data, 177
Spearman’s rank correlation, 147
Specificity, 74
Spider construction, 167
Splitting decision, 42
Splitting up data set, 71–74
SPSS, 153
Squashing, 49
Standardizing data, 24
Statistical performance, 9, 133
Stemming, 201
Stopping criterion, 45
Stopping decision, 42, 47

231

Stopword, 201
Supervised learning, 165
Support vector machines, 58–64
Support vectors, 60, 62
Support, 87, 89, 94–95
Suppression, 158
Survival analysis
evaluation, 117
measurements, 106–109
parametric, 111–114
semiparametric, 114–116
Survival function, 107
baseline, 116
System stability index (SSI), 136,
143
Swing clients, 170
Synonym, 178
T
Target
definition, 35–38
variable, 87
Test sample, 71
Test group, 170
Tie strength prediction, 203
Timeliness, 152
Time-varying covariates, 106, 116
Tool vendors, 7
Top decile lift, 76
Top-N recommendation, 183
Total data quality management
program, 152
Total quality management (TQM), 204
Traffic light indicator approach, 135,
137
Training sample, 45, 71
Training set, 51
Transaction identifier, 87
Transactional data, 14
Transform
logarithmic, 112
Trend analysis, 191
Triangle, 168
Truncation, 23
t-test, 143–144
Two-stage model, 52, 55
Types of data sources, 13–15
U
U-matrix, 101
Unary rating, 177
Undersampling, 166

232

▸ INDEX

Univariate
correlations, 29
outliers, 20
Universal approximation, 64
Universal approximators, 49
Unstructured data, 14
Unsupervised learning, 87, 100, 166
US Government Accountability Office,
156
Use limitation principle, 156
User agent, 186
User-based collaborative filtering, 176
User-item matrix, 177
V
Validation sample, 45
Validation set, 51
Validation
out-of-sample, 134
out-of-sample, out-of-time, 134
out-of-universe, 134
Value-added, 151
Vantage score, 146
Variable interactions, 32
Variable selection, 29–32
Vertex, 119
Virtual advisor, 184
Visit, 188
Visitors, 190
New, 190

Return, 190
Unique, 190
Visual data exploration, 17–19
W
W3C, 185
Weak classifier, 66
Web analytics, 4, 94, 185–195
Web beacon, 188
Web data collection, 185–188
Web KPI, 188–191
Web server log analysis, 185
Weight regularization, 51
Weighted average cost of capital,
37
Weights of evidence, 28–29
Weka, 153
White box model, 48
Wilcoxon test, 110
Winner take all learning, 70
Winsorizing, 23
Withdrawal inference, 16
Workflow net, 213
Y
Yahoo Search Marketing, 193
Z
z-score standardization, 24
z-scores, 22

Source Exif Data:

File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.7
Linearized                      : Yes
Author                          : Bart Baesens
Create Date                     : 2015:11:04 14:11:03Z
EBX PUBLISHER                   : John Wiley & Sons, Inc.
Modify Date                     : 2015:11:05 20:25:17Z
XMP Toolkit                     : Adobe XMP Core 5.4-c005 78.147326, 2012/08/23-13:03:03
Format                          : application/pdf
Creator                         : Bart Baesens
Title                           : Analytics in a Big Data World
Creator Tool                    : Adobe InDesign CS5.5 (7.5)
Metadata Date                   : 2015:11:05 20:25:17Z
Producer                        : Acrobat Distiller 9.4.0 (Macintosh)
Ebx Publisher                   : John Wiley & Sons, Inc.
Document ID                     : uuid:bc970f2b-dd01-4771-8c54-782ed2081e72
Instance ID                     : uuid:27d9ed91-707c-40fe-8ceb-590b2fc7c0e5
Page Layout                     : SinglePage
Page Mode                       : UseNone
Page Count                      : 252

EXIF Metadata provided by EXIF.tools

Analytics In A Big Data World Bart Baesens The Essential Guide To Science And Its Applicati

Bart%20Baesens%20Analytics%20in%20a%20Big%20Data%20World%20The%20Essential%20Guide%20to%20Data%20Science%20and%20its%20Applicati

Navigation menu

Versions of this User Manual:

Views

Navigation