Learning business rules with association rule classifiers

The main obstacles for a straightforward use of association rules as candidate business rules are the excessive number of rules discovered even on small datasets, and the fact that contradicting rules are generated. This paper shows that Association Rule Classification algorithms, such as CBA, solve both these problems, and provides a practical guide on using discovered rules in the Drools BRMS and on setting the ARC parameters. Experiments performed with modified CBA on several UCI datasets indicate that data coverage rule pruning keeps the number of rules manageable, while not adversely impacting the accuracy. The best results in terms of overall accuracy are obtained using minimum support and confidence thresholds. Disjunction between attribute values seem to provide a desirable balance between accuracy and rule count, while negated literals have not been found beneficial.