Mining Contrast Sets

1. Introduction

• Motivation
• understanding differences between groups
• Task
• provide an efficient algorithm for mining contrast contrast sets and pruning rules to reduce complexity
• provide post processing techniques to present subsets that are surprising
• control the false positives
• be statistically sound
• Goal
• To find contrast-sets whose support differs meaningfully (statistically) across groups
• $\exists i,j$  $P(cset = true | G_i) \neq P(cset = true | G_j) $, $max_{ij} | sup(cset, G_i) - sup(cset, G_j)| \geq min dev$

2. Naive Approach

• Add an attribute to the set (group type) and use Association Rule Mining to find the differences
• Problems
• this will not return group differences
• the results will be difficult to interpret