Using classification to evaluate the output of confidence-based association rule mining

Association rule mining is a data mining technique that reveals interesting relationships in a database. Existing approaches employ different parameters to search for interesting rules. This fact and the large number of rules make it difficult to compare the output of confidence-based association rule miners. This paper explores the use of classification performance as a metric for evaluating their output. Previous work on forming classifiers from association rules has focussed on accurate classification, whereas we concentrate on using the properties of the resulting classifiers as a basis for comparing confidence-based association rule learners. Therefore, we present experimental results on 12 UCI datasets showing that the quality of small rule sets generated by Apriori can be improved by using the predictive Apriori algorithm. We also show that CBA, the standard method for classification using association rules, is generally inferior to standard rule learners concerning both running time and size of rule sets.