A framework for evaluating knowledge-based interestingness of association rules

Abstract

In Knowledge Discovery in Databases (KDD)/Data Mining literature, "interestingness" measures are used to rank rules according to the "interest" a particular rule is expected to evoke. In this paper, we introduce an aspect of subjective interestingness called "item- relatedness". Relatedness is a consequence of relationships that exist between items in a domain. Association rules containing unrelated or weakly related items are interesting since the co-occurrence of such items is unexpected. 'Item-Relatedness' helps in ranking association rules on the basis of one kind of subjective unexpectedness. We identify three types of item-relatedness - captured in the structure of a "fuzzy taxonomy" (an extension of the classical concept hierarchy tree). An "item- relatedness" measure for describing relatedness between two items is developed by combining these three types. Efficacy of this measure is illustrated with the help of a sample taxonomy. We discuss three mechanisms for extending this measure from a two-item set to an association rule consisting of a set of more than two items. These mechanisms utilize the relatedness of item-pairs and other aspects of an association rule, namely its structure, distribution of items and item-pairs. We compare our approach with another method from recent literature.