Expert deduction rules in data mining with association rules: new results and challenges

Abstract

The paper deals with applications of domain knowledge in data mining with association rules. We consider enhanced association rules – general relations of general Boolean attributes derived from columns of analyzed data matrices. The challenge is to filter out rules – consequences of given items of domain knowledge. The general syntax of enhanced rules motivated the study of logical calculi whose formulas conform to these rules. Deduction rules in such calculi have proven inadequate for dealing with domain knowledge. This led to the definition and study of expert deduction rules – incorrect deduction rules, which are, however, supported by indisputable facts on the whole calculus. Such expert deduction rules can remarkably reduce the number of output association rules by omitting the consequences of given items of domain knowledge. This concerns both the association rules produced by the apriori algorithm and the enhanced association rules. However, expert deduction rules are available only for association rules evaluated using the confidence measure of interest. New expert deduction rules for association rules evaluated using the lift measure of interest are defined and studied. Two case studies demonstrate their usefulness. Finally, new challenges to study expert deduction rules are formulated.