Investigations into Relatedness-based Interestingness of Association Rules: A Transaction-driven Analysis

Abstract

An important problem in association rule (AR) mining is the identification of interesting ARs. In a retail market basket context, items may be related through various relationships like mutual interaction, 'substitutability' and 'complementarity'. We define them and present a classification of these relationships. We propose 'item-relatedness' of an item-pair as a composite of these relationships. We then present a structural decomposition of the relatedness of an item pair, based on its co-occurring transactions, co-occurring and non co-occurring item-neighborhoods. We identify those relationships that can be discerned solely from transaction data analysis. ARs that contain unrelated or weakly related item-pairs are likely to be interesting. The structural decomposition helps in clarifying components of relatedness. We finally analyze a typical scenario that contains objects revealing various shades of relatedness