Learning Statistically Significant Contrast Sets

Contrast set learning is important to discover control variables that can distinguish different groups in a dataset. Association rule mining has an inherent connection to the contrast set learning problem and has also been used to address it. All of the association rule based contrast set learning techniques use support-confidence based methods and inherit their limitations. In recent years statistically significant rule mining has become a viable alternative to address those limitations. We propose a novel contrast set learning approach based on statistically significant rule mining that eliminates the limitations in using traditional rule mining approaches and identifies statistically significant contrast sets. We evaluated our method by building a classifier using the discovered contrast sets. The performance of our classifier, while our method is not for classification per se, reveals the effectiveness of our approach in distinguishing the groups.