Clustering rules using empirical similarity of support sets

Abstract

We consider the problem of pruning a given set of if-then rules, such that the support of the pruned rule set is not much less than the support of the given rule set. An empirical measure of similarity between two rules is introduced. This similarity measure is proportional to the degree of overlap between the support sets of the two rules. Using this similarity measure, we cluster the given rule set via the complete linkage algorithm. Rules within a cluster are approximate substitutes for each other and, as such, they can be replaced by a single rule, which is chosen to be the rule whose individual support value is the largest in the cluster. The pruning procedure is demonstrated on a set of rules generated from a marketing data set.