Selection and aggregation of interestingness measures: a review

Abstract

Association Rule Mining is the process of retrieving frequent patterns that occur in a transaction database. Initially used as a market basket analysis solution for retail businesses, it has grown to cover many other fields such as medicine [1, 2], traffic estimation [3] and anomaly detection [4, 5]. An association rule has two components (antecedent and consequent) which is derived from a pattern (a set of items). However, it is not clear when investigating a frequent item set, which items imply the others (i.e., which is antecedent, and which is consequent). Therefore, several combinations of items as antecedent and consequent are generated. This leads to a huge amount of association rules being output by an algorithm for Association Rule Mining. Thus, it is imperative that data miners require some type of measures to evaluate the "interestingness" of these rules. There exist in excess of 70 well-known measures and countless other manually crafted measures in the literature. In this survey, we systematically discuss the methods which users could use to select or aggregate the interestingness measures, applicability of such methods and evaluation of the usage of such methods.