Master Thesis Project

Title Building a Global Model from Local Patterns by Pattern Team Discovery
Student Joris Valkonet
Supervisor Arno Knobbe
Abstract The discovery of patterns by rule induction algorithms like Subgroup Discovery, usually results in a large number of patterns. The patterns discovered tend to capture the essence of only a part of the dataset the algorithm was applied to. Unfortunately, the patterns are not mutually exclusive and, therefore contain a lot of redundancy. In order to capture the essence of the dataset with a small set of patterns, global models can be constructed from the patterns, so-called pattern teams. Pattern Team Discovery uses quality measures to qualify candidate pattern teams. An example of such a quality measure is the accuracy of a classifier learned on a candidate pattern team. The research in this thesis focuses on the performance of linear classifiers used to obtain the quality of a candidate pattern team. The second goal consists of whether a linear classifier performs better on a pattern team than non-linear classifiers. This is achieved by comparing the Decision Table Majority classifier with the Support Vector Machine classifier.