Inverted Heuristics in Subgroup Discovery

In rule learning, rules are typically induced in two phases, rule refinement and rule selection. It was recently argued that the usage of two separate heuristics for each phase—in particular using the so-called inverted heuristic in the refinement phase—produces longer rules with comparable classification accuracy. In this paper we test the utility of inverted heuristics in the context of subgroup discovery. For this purpose we developed a DoubleBeam subgroup discovery algorithm that allows for combining various heuristics for rule refinement and selection. The algorithm was experimentally evaluated on 20 UCI datasets using 10-fold double-loop cross validation. The experimental results suggest that a variant of the DoubleBeam algorithm using a specific combination of refinement and selection heuristics generates longer rules without compromising rule quality. However, the DoubleBeam algorithm using inverted heuristics does not outperform the standard CN2-SD and SD algorithms.