Organizing association rules with meta-rules using knowledge clustering

The effectiveness and the robustness of the existing association rule mining algorithms provides a huge number of high quality rules that the user can not understand. Consequently, thinking about another representation of the generated rules becomes vital task. For this, the present paper explores meta-rules discovery. It focuses on clustering association rules for large data sets. This allows the user better organizing and interpreting the rules. The main idea is to develop a clustering algorithm for the whole rules extracted by the association rule mining process (ARM). As a result, a set of meta-rules that the user or the data analyst can interpret are obtained, which permits to take a wise decision in a given domain. The issue for addressing such problem is to find an approach to perform clustering of association rules, which is a challenging problem not dealt with thus far. An adaptation of the k-means algorithm for association rules is proposed, by using new designed similarity measures and specific centroid computation. The clustering approach has been tested on a large data set obtained by merging three different public benchmarks. The result is promising. The proposed clustering generates three clusters, each includes one of the three benchmarks with a success rate varying from 70% to 90%.