Discriminatory Confidence Analysis in Pattern Mining

The field of association rule mining has long been dominated by algorithms that search for patterns based on their frequency of occurrence in a given dataset. The birth of weighted association rule mining caused a fundamental paradigm shift in the way that patterns are identified. Consideration was given to the "importance" of an item in addition to its frequency of occurrence. In this research we propose a novel measure which we term Discriminatory Confidence that identifies the extent to which a given item can segment a dataset in a meaningful manner. We devise an efficient algorithm which is driven by an Information Scoring model that identifies items with high discriminatory power. We compare our results with the classical approach to association rule mining and show that the Information Scoring model produces widely divergent results. Our research reveals that mining on the basis of frequency alone tends to exclude some of the most informative patterns that are discovered using discriminatory power.