Algorithms for filtration of unordered sets of regression rules

Abstract

This paper presents six filtration algorithms for the pruning of the unordered sets of regression rules. Three of these algorithms aim at the elimination of the rules which cover similar subsets of examples, whereas the other three ones aim at the optimization of the rule sets according to the prediction accuracy. The effectiveness of the filtration algorithms was empirically tested for 5 different rule learning heuristics on 35 benchmark datasets. The results show that, depending on the filtration algorithm, the reduction of the number of rules fluctuates on average between 10% and 50% and in most cases it does not cause statistically significant degradation in the accuracy of predictions. © 2012 Springer-Verlag.