Outlier Detection
and
Explanation
Bachelor‘s thesis defence
Václav Blahut, 395963
June 2015
Outliers
• Objects that do not comply with
general behavior or model of the
data
• Many methods of outlier
detection
• Few methods for class outlier
detection, outlier explanation or
non-singleton outlier detection
• Many applications (fraud
detection, system health
issues…)
Class outlier
Class outlier
General outlier
Non-singleton
outlier
Examples of outliers
About this method
• Proposed by Angiulli and Fassetti in 2009
• Unsupervised, ILP-based
• Outlier is a set of examples
• Datasets in Aleph format
• Strongly relational
• Background knowledge – determinations, modes, facts
• Positive and negative examples
The Algorithm
• Main idea – excluding outlier from dataset will simplify its model
• Compares models (Aleph theories) generated including and excluding
each possible outlier
• Three types of abnormality
• Irregularity
• Anomaly
• Outlier
• Implemented in Prolog
• Output – set of detected abnormalities with their explanations
Evaluation
• Four datasets
• ZOO
• Student Loan 100
• Small and Medium Enterprises
• House Votes 84 (Republicans vs. Democrats)
• Results compared to weka-peka and weka-CODB
• Explanations compared to RF-OEX
Thank you for your attention.