Understandability of association rules: A heuristic measure to enhance rule quality

Abstract

Association Rule (AR) mining has been plagued by the problem of rule immensity. The sheer numbers of discovered rules render comprehension difficult, if not impossible. In addition, due to their representing commonplace facts, most of these rules do not add to the knowledge base of the user examining them. Clustering is a possible approach that mitigates this rule immensity problem. Clustering ARs into groups of 'similar' rules is advantageous in two ways. Related rules being placed together in the same cluster facilitates easy exploration of connections among the rules. Secondly, a user needs to examine only those rules in 'relevant' clusters. The notion of weakness of an AR is introduced. Weakness reveals the extent of an AR's inability to explain the presence of its constituents in the database of transactions. We elaborate on its usefulness and relevance in the context of a retail market. After providing the intuition, a distance-function on the basis of weakness is developed. This distance-function forms the basis of clustering ARs. Average linkage method is used to cluster ARs obtained from a small artificial data set. Clusters thus obtained are compared with those obtained by applying a commonly used method (from recent data mining literature) to the same data set. © Springer-Verlag Berlin Heidelberg 2007.