Mining Association Rules for Estimation and Prediction

The standard Basket Analysis derives all frequent itemsets and all association rules having support and con dence levels greater than their thresholds, and lters out trivial rules in statistical sense (1994, 1991). This framework gives comprehensive "descriptions" of regularities contained in the data. Another major purpose of data mining is to derive important knowledge for "estimation and prediction" on the underlying system which has generated the data (1996). We propose a novel principle to derive association rules for the latter purpose, where the rules provide maximal guesses from minimal facts about the system while maintaining their support and con dence levels as uniform as possible. The principle and its evaluation through real world data are described in the later sections.