Efficient Statistical Pruning of Association Rules
[Proceedings of the European Conference on Machine Learning]
Association mining is the comprehensive identification of frequent patterns in discrete tabular data. The result of association mining can be a listing of hundreds to millions of patterns, of which few are likely of interest. In this paper we present a probabilistic metric to filter association rules that can help highlight the important structure in the data. The proposed filtering technique can be combined with maximal association mining algorithms or heuristic association mining algorithms to more efficiently search for interesting association rules with lower support.
Association mining is the process of identifying frequent patterns in a tabular dataset, usually requiring some minimum support , or frequency of the pattern in the data . The discovery of frequent patterns in the data is usually followed by the construction of association rules , which portray the patterns as predictive re- lationships between particular attribute values. Unfortunately, since association mining is an exhaustive approach, it is possible to generate many more patterns than a user can reasonably evaluate. Furthermore, many of these patterns may be redundant. Thus, is it important to develop informed and efficient pruning systems for association mining rules.
A number of methods to filter association mining results have been published, and they can be classified along several lines. First, the filtering can be objective or subjective . Our goal is to design an objective, or purely computational, filter, both for inter-discipline generality and to avoid the bias introduced by subjective evaluation. In a separate categorization, rule filtering can be done on a rule-by-rule basis , or in an incremental manner where rules deemed in- teresting are gradually added to a rule list set  or probability model . We focus our attention on the rule-by-rule approach. This approach is appropriate in application areas where dense data tables lead to the generation of millions of as- sociation rules, a set too large to use directly in incremental filtering algorithms. As an added advantage, filtering rules independently allows the straightforward use of batch parallelization of the filtering to produce linear speed-ups.