The problem of mining association rules has attracted a lot of attention in the research community. Several techniques of efficient discovery of association rules have appeared. Those rules which exceeded a predetermined minimum threshold for support and confidence are considered to be interesting. However, association rule mining does not discover the true correlation relationship, because high minimum support usually generates common sense knowledge, while low minimum support generates huge number of rules, the majority of which are uninformative. Therefore, many metrics of interestingness, such as Convictional Loevinger, Centered Confidence, Lift, etc… have been devised to help find interesting rules while filtering out uninteresting ones. The application of those measures left the user in front of several shortcomings due to the impossibility of filtering out uninteresting rules and respectively removing interesting ones. Those problems are directly linked to the choice of couplets (metric, threshold). The use of metrics is meant to reduce the number of rules. However, after their filtration, the number is still huge which confronts the domain expert with problems during validation. Our approach consists in attributing to every rule its own vector of metrics, then clustering rules into groups relying on the vectors of measures’ value. Besides, every generated cluster is described by the vector of its centre. From this viewpoint, a domain expert can simply validate the centers of those clusters. The result of the expert’s decision can be automatically generalized to all members (rules) of the clusters. This approach allows us to reduce significantly the cognitive effort provided in this task.

Association Rules, Interestingness, Metrics, Clustering.