Anomaly Labeling Instructions

User Manual: Pdf

Open the PDF directly: View PDF PDF.
Page Count: 2

DownloadAnomaly Labeling Instructions
Open PDF In BrowserView PDF
Anomaly Labeling Instructions
The objective is to indicate the timestamps at which an anomaly is identified to
start, according to the instructions below. Labels should not be done for data files
in the directory nab/data/realKnownCause.
1. The first 15% of a data set is the probationary period. Do not label any
anomalies inside the probationary period, but do observe the data patterns
inside the probationary period – these patterns are learned as normal so if you
see these patterns again in the rest of the data, they are not anomalies; that is
unless a new learned normal occurs. In the data visualizer, the probationary
period has been shaded grey.
2. If the normal pattern changes in the data (after the first 15%), the first data
point of the change is the start of an anomaly – e.g. point “a” in Fig. 1 below.
A new pattern in the data is established as the new normal if it proceeds for
approximately 10% of the data file length – e.g. at point “b” in Fig. 1 below,
the new pattern, a straight line, has been learned.
3. If after a new normal is established, the data stream goes back to a previously
learned pattern – i.e. a remembered normal – then this transition is not an
anomaly. E.g. in Fig. 1 below, the original data pattern resumes at point “c”;
this is a remembered normal and should not be labeled as an anomaly.
4. A long flat line becomes an anomaly if the period of flatness exceeds the
longest period that was in the normal data before that point. For instance, in
Fig. 2 below, some periods of the data at zero are learned in the probationary
period; the green arrow on the left shows the longest zero-period that is
learned. When another, longer zero-period occurs after the probationary
period is over (the green arrow on the right), the anomaly should be labeled
starting after the length of the previous longest learned period, at point “a”.
5. A significantly different data value – i.e. more than 20% different – can be
labeled an anomaly, even if the value isn’t a new maximum or minimum
value. For instance, in a normal pattern of spikes that reach values of 10 and
60, a new spike that reaches 30 is an anomaly, even though this spike’s value
is between the normal peaks.
6. Don’t over analyze trying to locate anomalies. In general there should be very
few anomalies per data file. One test to use: eliminate all future data from
your thinking and ask yourself, “Assuming this is an important streaming
data source, would you wake someone up if this occurred?"
7. If there is a lot of noise in the data, a pattern change has to be more obvious to
be an anomaly. In very clean data, almost any new value would be
anomalous; in noisy data it isn’t.
8. Ignore all “future” data as you move through a dataset looking for anomalies.
That is, try to identify the start of an anomaly in real-time, as a detector
would.

9. If in doubt whether there are one or two anomalies in a relatively small
window, merge two anomalies into one.
10. Save the labels file as a JSON with naming convention ‘XY_labels_vAB.json’,
where ‘XY’ are your initials and ‘vA.B’ refers to the corresponding NAB
version number. The format should be a dictionary of key-value pairs for
each data file; the key is the data file location, and the value is a list of
timestamps indicating the anomaly start times. The script nab/scripts/
create_empty_label_file.py will create an empty JSON file for labels, and for
an example please see nab/labels/raw/AL_labels_v1.0.json.
11. Keep in mind that anomalies are rare by definition, and this is an underlying
assumption of NAB parameters and logic.
12. Anomalous data points within close proximity are assumed to represent the
same anomalous data, so please choose one of the data points as your label.
Here we define close proximity as 1% the length of a data file. If your raw
labels contain timestamps that are within close proximity, an error will be
raised that requests you fix the labels.

Fig. 1

Fig. 2
Source Exif Data: 
File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.3
Linearized                      : No
Tagged PDF                      : Yes
Language                        : en-US
Page Count                      : 2
XMP Toolkit                     : 3.1-702
About                           : 15F2EA4C-0DF7-3FC6-84F8-AE889135B0D2
Keywords                        : 
Producer                        : http://www.convertapi.com
Modify Date                     : 2015:10:14 17:10:39-06:00
Create Date                     : 2015:10:14 17:10:39-06:00
Metadata Date                   : 2015:10:14 17:10:39-06:00
Creator Tool                    : Microsoft® Word 2013
Format                          : application/pdf
Description                     : 
Creator                         : Celeste Baranski
Title                           : 
Document ID                     : uuid:FBE99229-A79A-650B-9399-BBEBE7734275
Instance ID                     : uuid:15F2EA4C-0DF7-3FC6-84F8-AE889135B0D2
Author                          : Celeste Baranski
EXIF Metadata provided by EXIF.tools

Navigation menu