Mining Multiple Level Non-redundant Association Rules through Two-Fold Pruning of Redundancies

Association rules (AR) are a class of patterns which describe regularities in a set of transactions. When items of transactions are organized in a taxonomy, AR can be associated with a level of the taxonomy since they contain only items at that level. A drawback of multiple level AR mining is represented by the generation of redundant rules which do not add further information to that expressed by other rules. In this paper, a method for the discovery of non-redundant multiple level AR is proposed. It follows the usual two-stepped procedure for AR mining and it prunes redundancies in each step. In the first step, redundancies are removed by resorting to the notion of multiple level closed frequent itemsets, while in the second step, pruning is based on an extension of the notion of minimal rules. The proposed technique has been applied to a real case of analysis of textual data. An empirical comparison with the Apriori algorithm proves the advantages of the proposed method in terms of both time-performance and redundancy reduction.