My CBR Guide
User Manual:
Open the PDF directly: View PDF 
.
Page Count: 12
Explanation Capabilities of the Open Source
Case-Based Reasoning Tool myCBR
Thomas R. Roth-Berghofer1,2and Daniel Bahls1,2
1Knowledge Management Department,
German Research Center for Artificial Intelligence DFKI GmbH
Trippstadter Straße 122, 67663 Kaiserslautern, Germany
2Knowledge-Based Systems Group, Department of Computer Science,
University of Kaiserslautern, P.O. Box 3049, 67653 Kaiserslautern
{thomas.roth-berghofer,daniel.bahls}@dfki.de
Abstract. This paper describes the various explanation capabilities of
the open source case-based reasoning tool myCBR.myCBR features con-
ceptual explanations, which provide information about concepts of the
application domain, backward explanations, which explain results of the
retrieval process, and forward explanations, which support in the mod-
elling of similarity measures. myCBR has been developed as a rapid pro-
totyping tool with a general purpose interface as well as a similarity-based
retrieval engine for easy integration in other applications where the ex-
planations can be further adapted to the application’s requirements.
Key words: Case-Based Reasoning, explanation
1 Introduction
Ease-of-use as well as approachability of any software system is improved by
increasing its understandability, which in turn can be supported by appropri-
ate explanation capabilities [1, 2]. We follow Schank [3] in considering expla-
nations the most common method used by humans to support understanding
and decision making. In everyday human-human interactions explanations are
an important vehicle to convey information in order to understand one another.
Explanations enhance the knowledge of the communication partners in such a
way that they accept certain statements. They understand more, allowing them
to make informed decisions.
This communication-oriented view leads to the following explanation scenario
with the three participants (Figure 1) originator,user, and explainer. The orig-
inator provides something to be explained, e.g., the solution to some problem, a
technical device, a plan, etc. Here, the originator comprises the modelling tools
and the retrieval engines of myCBR. The user is the addressee of the explana-
tion, and the explainer provides the explanation. The explainer is interested in
transferring the intention of the originator to the user as correctly as possible.
The explainer chooses the kind of the explanation [4] and is responsible for the
Fig. 1. Participants in explanation scenario [2]
computational aspects of the explanation process. The originator and explainer
need to work together rather tightly to improve the communication with the
user. The originator needs to provide the appropriate information in order to
allow the explainer to construct appropriate explanations.
In order to make use of an information system the user needs at least a basic
understanding of the application domain, i.e., the respective terms and concepts.
But usually the user is not familiar with all of them. Conceptual explanations
provide information about concepts of the application domain, linking unknown
concepts to already known concepts.
In order to support the communication scenario described above, myCBR
provides two general kinds of explanations: forward and backward explanations.
Forward explanations explain indirectly, presenting different ways of optimising
a given result and opening up possibilities for the exploratory use of a device
or application. Backward explanations explain the results of a process and how
they were generated.
The rest of the paper is structured as follows: After a brief introduction of
myCBR in Section 2 the supported kinds of questions are presented in Section 3
followed by corresponding explanations (Section 4). We then describe the inte-
gration of the explanations into the user interface of Prot´eg´e in Section 5 and
the accessible explanation data structures in Section 6. We close the paper with
a summary and outlook.
2 The Open-Source Case-Based Reasoning Tool myCBR
myCBR3is an open-source plug-in for the open-source ontology editor Prot´e-
g´e4. Prot´eg´e [5] allows to define classes and attributes in an object-oriented way.
Furthermore, it manages instances of these classes, which myCBR interprets as
cases. So the handling of vocabulary and case base is already provided by Pro-
t´eg´e. The myCBR plug-in provides several editors to define similarity measures
for an ontology and a retrieval interface for testing.
The main motivation for the development of myCBR was the need for a com-
pact and easy-to-use tool for building prototype CBR applications in teaching,
research, and small industrial projects with minimal effort [6]. The tool needed
3http://mycbr-project.net
4http://protege.stanford.edu/
to be easily extendable to allow the experimental evaluation of new algorithms
and recent research results. Many ideas for the implementation of myCBR came
from CBR-Works5[7] which is not supported any more.
The current version of myCBR focuses on the similarity-based retrieval step
of the CBR cycle [8]. A popular example of such retrieval-only systems are
case-based product recommender systems [9]. While the first CBR systems were
often based on simple distance metrics, today many CBR applications make use
of highly sophisticated, knowledge-intensive similarity measures [10].
As the main goal of myCBR is to minimise the effort for building CBR appli-
cations that require knowledge-intensive similarity measures, myCBR provides
comfortable graphical user interfaces for modelling various kinds of attribute-
specific similarity measures and for evaluating the resulting retrieval quality. In
order to reduce also the effort of the preceding step of defining an appropri-
ate case representation, it includes tools for generating the case representation
automatically from existing raw data.
myCBR provides retrieval mechanisms to find similar cases for a specified
query. Both functionalities, modelling and retrieval, are available in separate
tabs of the Prot´eg´e editor. Eventually, a myCBR model can be integrated into
other applications, for which a standalone API is provided. In addition, since a
CBR model is used as a background of an application we will later present a
third tab, which is used to define conceptual explanations (Section 6.1).
From its conception, myCBR was designed with improved communication
between the system and the user—knowledge engineer and end-user—in mind.
The novice as well as the expert knowledge engineer is supported during the
development phase of a myCBR project through intelligent support approaches
and advanced GUI functionality. A dedicated explanation component provides
modelling support information as well as explanations of retrieval results for
quicker round trips of designing and testing.
3 Questions for myCBR
Explanations are in principle answers to questions. Hence, we formulated the
requirements for the explanation support in terms of questions for which we
developed explanation schemes (see Section 4). From the many questions, some of
which are interesting at modelling time and others at retrieval time, we selected
the most often asked questions, i.e., about concepts, about retrieval outcomes,
and questions arising during modelling and maintenance.
3.1 Questions about Concepts
Questions about terms and concepts often arise for the end user when he or she
is not familiar with the application domain. The user must be familiar with the
5CBR-Works was developed at the University of Kaiserslautern in co-operation with
empolis GmbH, formerly tecinno GmbH.
terms and concepts used in the application in order to take advantage of it. The
respective question is very simple:
a) What is meant by this concept?
Fortunately, this kind of questions can be also answered quite easily, even though
additional knowledge is needed.
3.2 Questions about a Retrieval Outcome
A user may be surprised by a particular case’s similarity value or she wants
to assure herself of the system’s quality. Both situations offer opportunities to
increase trust into the system.
b) How did the system come to the similarity assessment of a particular case?
c) Which are the most similar aspects of a case? Which are the least?
d) Why are some demands of the query not met by the most similar cases?
The first two questions come to mind when a particular case is under examina-
tion. Asking question b) the user wants to retrace the procedure of similarity
assessment. Especially when the outcome was surprising, the answer may de-
liver a justification or reveal a modelling error within the similarity measure. An
answer to question c) explains in which way the case is similar to the query.
Another issue which rather concerns the whole ranking than a particular
case is the absence of a good solution that meets the constraints in the query
satisfyingly (addressed by question d). The user is unhappy with the top ranked
cases of the retrieval result. If the estimated similarity of the case does not
comply with the utility for the user, the similarity measure is insufficient. The
source can then be found by asking questions of type b). If the similarity measure
is correct though, a good case is missing in the case base, and the top cases are
the best available.
3.3 Questions during Modelling and Maintenance
Forward explanations are intended to assist the knowledge engineer during mod-
elling time, i.e., regarding the knowledge containers vocabulary and case base [11].
They concern the system’s future behaviour. This situation here requires an anal-
ysis of the system as a whole in its current configuration.
e) Are some problem types underrepresented in the case base?
f) Is there an imbalance of cases in the case base?
g) Which parts of the similarity measure are of high or low relevance?
h) Which symbols are similar to a given symbol?
The first two questions address related issues. The case base may contain cases
which appear to be very much the same while some special cases seem to be
missing. Imagine a doctor who medicates patients having a cold a lot more
often than patients having malaria. He does not need to remember every patient
who had a cold. But probably it is useful to remember every patient who had
malaria. On the other hand, regarding the used cars domain, it is probably not
helpful, if the explainer states that the model lacks cars having a price lower than
1,000 Euro and being manufactured one year ago. In general, the attributes of
a model are not statistically independent.
The knowledge engineer tries to approximate the utility of a case with the
help of similarity measures. This can be a quite complex task, and getting lost
in details can happen quickly. We want to guide the user to the relevant aspects
of a similarity measure by answering question g).
The last question addresses the local similarity measure for attributes of
symbolic type. The size of a symbol table grows quadratically with the number
of allowed values. In this case, the task of keeping this table consistent during
maintenance work becomes hard.
4myCBR’s Explanations
Even though the principles of case-based reasoning are easy to understand some-
times a particular retrieval result is not. We distinguish between the explanations
for knowledge engineers, which are supposed to assist in modelling and main-
taining, and the explanations for the end user, who wants to understand the
system’s behaviour and the concepts involved.
If the knowledge engineer comes to the conclusion that he encountered a
modelling error, several action alternatives are available. He or she can insert,
delete or modify existing cases, modify similarity measures or even change the
current vocabulary for which each task offers again a variety of alternatives in
detail and needs a lot of attention [12]. The choice of action must be founded on
the analysis of the knowledge base with respect to the purpose of the application.
In such a situation, the explanatory capabilities of a CBR system are certainly
limited, because it is not provided with a deeper understanding of the application
domain. However, it can give hints, illustrate internal dependencies and highlight
certain parts of the model in such a way that the knowledge engineer is enabled
to make his decision faster and more confidently.
In the context of case-based reasoning, conceptual explanations are used to
explain the vocabulary knowledge container. Backward explanations explain the
outcome of a particular retrieval result and provide means for understanding
the results of a similarity calculation [13]. Forward explanations are intended
for modelling and maintenance assistance, by providing information about the
status of the model.
An explanation scheme, here, describes a special procedure to answer a par-
ticular kind of question. During the design of an explanation scheme we kept in
mind that for an explanation to count as good it needs to be relevant,innova-
tive,compact,correct, and convincing [14]. We do not consider innovativeness
here, because it requires some kind of user model and some facility that records
when an explanation has been provided. We intend to use the justification-based
explanation support server Reduxexp in the future [15].
4.1 Conceptual Explanations
A conceptual explanation is a comprehensive description of a concept. It con-
sists of a definition, some examples and references to further characterisations,
for which any kind of medium can be used (e.g., text, images, audio, video).
Conceptual explanations are inherently static, because concepts usually do not
change. But there are good reasons to consider the context in which the concept
is used and the user’s personal level of knowledge. However, in this work we
regard a conceptual explanation as static regarding the given ontology.
We want to support the end user and the knowledge engineer. For the end
user knowledge about the concepts is the most important conceptual knowledge.
This kind of knowledge is domain dependent and not yet part of a myCBR model.
It must be additionally supplied by the knowledge engineer.
An important source for conceptual explanations for the end user is the vo-
cabulary knowledge container of a CBR system [16]. CBR systems do not provide
sufficient knowledge to provide satisfying conceptual explanations per se. This
also applies for myCBR in its role as the originator in the explanation scenario
(see Figure 1). Thus, the vocabulary knowledge container, i.e., the ontology,
needs an extension.
Conceptual explanations for the knowledge engineers are about concepts and
functionalities of myCBR itself, for which help buttons and other documentation
are already provided.
4.2 Backward Explanations
Question b) will be answered with the help of a detailed recording of the retrieval
process. All local and global similarity measures involved write comments and
provisional similarity values to a certain protocol. So, every step of similarity
calculation can be traced.
In order to answer Questions c) and d) the notion of aspect must be clarified.
Here, an aspect of a query is one single attribute. The similarity of an aspect is
then a local similarity value and can be found in the mentioned protocol. Thus,
sorting the attributes of a case with respect to their (local) similarity values
gives the answer to question c).
4.3 Forward Explanations
Both questions e) and f) ask for some kind of distribution within the case base.
For each attribute the value distribution within the case base is used to show the
user which attribute values have many representations and which have only a
few. In a similar way, a distribution of class instances is set up. Even if the value
distribution was uniform for every class and attribute of the vocabulary, the
case distribution may not be uniform at all due to interdependencies between
attributes. On the other hand, if the value distribution is irregular for some
attributes or classes, the case distribution cannot be uniform at the same time.
Hence, the questions are only partly answered.
Similarity measures provide a means to compare two objects of a certain
domain with each other. Since some comparisons are more frequent than others,
the corresponding part of the similarity measure is also of higher relevance than
other parts. To answer Question g), the task is to find out the comparison fre-
quency for two arbitrary objects for each attribute domain. This will be referred
to as a relevance distribution in the following. On the one hand, this depends on
the value distribution within the case base. On the other hand, it depends on the
value distribution within the submitted queries by the user. This piece of infor-
mation is not given, because it involves the continuous analysis of user queries6.
For now we provide a preliminary solution by assuming the value distribution
within the case base to be similar to the one within the user queries.
The average number of comparisons between a query value qand a case value
cis given by the following formula:
freqrelevance(q, c) = freqcasebase(q)∗freqcasebase(c)
where freqcasebase(x) is the number of cases having the attribute value x divided
by the size of the case base.
A look at the local similarity measure gives the answer to question h). But
keeping track of the contents of a big table becomes more difficult as the number
of allowed symbols increases. Trying to understand the origin of this problem,
one realises that it cannot only follow from its quadratic size. Another issue is
the chaotic arrangement of the symbols regarding the users varying exploration
interest. Because the row and column header are in a fixed symbol order, the
comparison of some relevant similarities demands high concentration and a good
eye. To solve this problem we introduced the concept of dynamic symbol orders
where symbols are ordered by ascending or descending similarity in a selected
row or column.
5 User Interface Integration
As ease-of-use is of high priority for the development of myCBR we aimed at
seamlessly integrating explanations into the user interface.
Conceptual and backward Explanations In order to increase transparency and
trust in the retrieval process [16], myCBR creates an explanation object for each
case during similarity calculation.
Figure 2 gives a schematic overview of the Retrieval GUI. The area is divided
into several columns. The leftmost column is used for query specification. The
others are used to show the retrieval results. The rightmost column lists all cases
of the case base ordered by their similarity to the query. The number of columns
displayed in between can be configured. The rows of the table are labelled with
the names of the classes’ attributes.
6Collecting and analysing user queries is an important maintenance and quality im-
provement task (cf. [12])
Fig. 2. Schematic overview of retrieval GUI
Conceptual explanations are addressing the end users of the system. They are
interested in retrieving the most similar cases to their queries. At the bottom
of the retrieval GUI conceptual explanations are shown when the user hovers
the mouse over a table cell. Figure 3 shows a snapshot. Retrieval details are
Fig. 3. Retrieval interface with conceptual and backward explanations
presented to the user either as tool tips or in abbreviated form along with the
case’s attribute value, e.g., the mileage of car offer 561 audi (18,940) is 100%
similar to the requested mileage (20,000). Another valuable feature is the option
to find the most similar cases with respect to a single attribute by simply clicking
on the attribute name (row head). In attribute rich cases one might also want to
sort the local similarity values of one case. For this one clicks on the respective
case name (column head).
Fig. 4. Schematic overview of similarity measure editing with explanation panel
Forward Explanations While developing a CBR system an important question
is whether a similarity measure leads to the appropriate cases for a given query.
We distinguish between local and global similarity measures. This is reflected in
the GUI (Figure 4). Depending on whether a class or slot is selected, a global
or local similarity measure editor is displayed on the right side of the window.
A corresponding explanation panel is superimposed on demand.
6 Implementation Issues
The support of explanations requires new components and changes of the cur-
rent myCBR system. As we mentioned earlier, the explainer must be strongly
interlaced with the originator due to reliability and authenticity issues (Fig-
ure 1). The explainer needs an insight into the originator’s knowledge, and the
originator must actively deliver information about its behaviour to the explainer.
Fig. 5. Detailed components of originator and explainer
Considering the different kinds of explanation, one notices that a central and
easily accessible explanation component, which we named explanation manager,
is needed to generate conceptual and forward explanations. The explanation
manager also provides backward explanations using logged similarity calcula-
tions. Figure 5 gives an overview of the GUI extensions of the similarity measure
editors and the retrieval widget as well as the additional editor adding conceptual
knowledge to a myCBR model.
6.1 Implementation of Conceptual Explanations
Conceptual explanations do not involve complicated algorithms. The required
functionality at its core is a static mapping from concepts to explanations.
Figure 6 shows the conceptual explanations editor. On the left hand side a tree
Fig. 6. Editor for conceptual explanations
displays all concepts of the ontology. On the right hand side one can edit its
explanation, i.e., a short description and a list of URLs of further documents.
For example, manufacturers in the used cars application can be explained via
wikipedia7or google define8.
The conceptual explanations are stored as part of the myCBR model, but
in a separate file in order to separate it from the reasoning knowledge. When a
myCBR model is loaded, the explanation manager looks for this file.
7http://wikipedia.org
8http://www.google.com/help/features.html#definition
6.2 Implementation of Backward Explanations
Backward explanations are given with the help of a certain kind of retrieval
recording. Since we want to log the calculation steps of each global and local sim-
ilarity measure, the interface float ←SIM(query, case) which is implemented
by all similarity measures is extended to float ←SIM(query, case, explanation).
Every similarity measure was modified in order to write the resulting similar-
ity value and further comments such as intermediate results to the explanation
object while the similarity between query and case is determined. A tree-like
structure, mirroring the local-global structure of similarity measures, is used to
build the overall backward explanation for a case. The query and case values as
well as the responsible similarity measure are tracked also. The retrieval engine
is responsible for connecting them appropriately.
6.3 Implementation of Forward Explanations
Statistical information about the value distribution within the case base is re-
quired to give forward explanations. Furthermore, this information cannot be
gained in tow of another procedure as it was the case for the backward ex-
planations. The case base must be examined in an active way. Certainly, it is
not advisable to let each explanatory component calculate its required statistics
on its own as soon as an explanation is demanded, because the computation
time for each given explanation would be annoyingly high. We decided to use
the explanation manager as an intermediate component that holds all statistical
information about the case base.
7 Summary
Three different kinds of explanation are delivered by myCBR:Conceptual expla-
nations describe items of the vocabulary knowledge container to the end user
via textual descriptions or via links to additional information, e.g., documents.
Backward explanations explain the retrieval outcome in relation to a particular
query to the end user. And, forward explanations assist the knowledge engineer
during modelling and maintenance. We formulated questions and corresponding
explanation schemes and extended the user interfaces of myCBR accordingly to
support the explanation presentation.
myCBR is still an ongoing project. Several extensions of the system are
planned or are already under development. We encourage other researchers to
try out myCBR in their own research and teaching projects and to contribute to
the further development by implementing their own extensions and experimental
modules.
References
1. Roth-Berghofer, T.R.: Explanations and Case-Based Reasoning: Foundational is-
sues. In Funk, P., Gonz´alez-Calero, P.A., eds.: Advances in Case-Based Reasoning,
Springer-Verlag (2004) 389–403
2. Roth-Berghofer, T.R., Richter, M.M.: On explanation. K¨unstliche Intelligenz 22(2)
(2008) 5–7
3. Schank, R.C.: Explanation Patterns: Understanding Mechanically and Creatively.
Lawrence Erlbaum Associates, Hillsdale, NJ (1986)
4. Roth-Berghofer, T., Cassens, J., Sørmo, F.: Goals and kinds of explanations in
case-based reasoning. In Althoff, K.D., Dengel, A., Bergmann, R., Nick, M.,
Roth-Berghofer, T., eds.: WM 2005: Professional Knowledge Management, Kaisers-
lautern, Germany, DFKI GmbH (2005) 264–268
5. Gennari, J.H., Musen, M.A., Fergerson, R.W., Grosso, W.E., Crub´ezy, M., Eriks-
son, H., Noy, N.F., Tu, S.W.: The evolution of Prot´eg´e an environment for
knowledge-based systems development. Int. J. Hum.-Comput. Stud. 58(1) (2003)
89–123
6. Stahl, A., Roth-Berghofer, T.R.: Rapid prototyping of CBR applications with
the open source tool myCBR. In Bergmann, R., Althoff, K.D., eds.: Advances in
Case-Based Reasoning, Springer Verlag (2008)
7. Schulz, S.: CBR-Works: A state-of-the-art shell for case-based application building.
In Melis, E., ed.: Proceedings of the 7th German Workshop on Case-Based Rea-
soning, GWCBR’99, W¨urzburg, Germany, University of W¨urzburg (1999) 166–175
8. Aamodt, A.: Explanation-driven case-based reasoning. In Stefan Wess, K.D.A.,
Richter, M., eds.: Topics in Case-Based Reasoning, Berlin, Springer-Verlag (1994)
9. Bridge, D., G¨oker, M.H., McGinty, L., Smyth, B.: Case-based recommender sys-
tems. Knowledge Engineering Review 20(3) (2006)
10. Stahl, A.: Learning of Knowledge-Intensive Similarity Measures in Case-Based
Reasoning. PhD thesis, University of Kaiserslautern (2003)
11. Richter, M.M.: The knowledge contained in similarity measures. Invited Talk at
the First International Conference on Case-Based Reasoning, ICCBR’95, Sesimbra,
Portugal (1995)
12. Roth-Berghofer, T.R.: Knowledge Maintenance of Case-Based Reasoning Systems
– The SIAM Methodology. Volume 262 of Dissertationen zur K¨unstlichen Intelli-
genz. Akademische Verlagsgesellschaft Aka GmbH / IOS Press, Berlin, Germany
(2003)
13. Bahls, D., Roth-Berghofer, T.: Explanation support for the case-based reasoning
tool myCBR. In: Proceedings of the Twenty-Second AAAI Conference on Artificial
Intelligence. July 22–26, 2007, Vancouver, British Columbia, Canada., The AAAI
Press, Menlo Park, California (2007) 1844–1845
14. Swartout, W.R., Moore, J.D.: Explanation in second generation expert systems.
In David, J., Krivine, J., Simmons, R., eds.: Second Generation Expert Systems,
Berlin, Springer Verlag (1993) 543–585
15. Roth-Berghofer, T.R., Mittag, F.: ReduxExp: A justification-based explanation-
support server. Proceedings of AI-2008. The twenty-eighth SGAI international con-
ference on artificial intelligence. In Petridis, M., Coenen, F., Bramer, M., eds.:
Research and Development in Intelligent Systems XXV, London, UK, Springer
Verlag (2008)
16. Roth-Berghofer, T.R., Cassens, J.: Mapping goals and kinds of explanations to
the knowledge containers of case-based reasoning systems. In Mu˜noz-Avila, H.,
Ricci, F., eds.: Case-Based Reasoning Research and Development, 6th International
Conference on Case-Based Reasoning, ICCBR 2005, Chicago, IL, USA, August
2005, Proceedings. Number 3620 in Lecture Notes in Artificial Intelligence LNAI,
Heidelberg, Springer Verlag (2005) 451–464
