Novel Interestingness Measures for Mining Significant Association Rules from Imbalanced Data

Abstract

Associative classification is a rule-based approach that joins Association Rule Mining and Classification to build classifiers that predict class labels for new data. Associative classifiers may generate an overwhelming number of rules which are hard to handle. Delving through these rules to identify the most interesting ones is a challenging task. To overcome this problem, several measures have been proposed. However, for imbalanced datasets, existing measures are no more reliable. In fact, they tend either to favour rules of major classes and consider others as uninteresting or only emphasize on the rules of minor classes and omit other ones. In this respect, we propose five new measures which tend to be fair for both types of classes regardless of their imbalanced distribution. Extensive carried out experiments on real-world datasets show that the new measures are able to efficiently extract significant knowledge from minor classes without decreasing the global predictive accuracy.