Preventing overfitting of decision rule induction using discernibility based clustering

Abstract

The authors have proposed a rule induction method using discernibility clustering. In this method, we apply clustering to the set of the positive objects using a similarity measure reflecting discerni- bility from the set of negative objects. Comparing LEM2, which is a conventional rule induction method, the proposed method obtained simpler rule sets, but it tends to be worse with respect to classification accuracy. In this paper, we improve clustering accuracy of the proposed method by preventing overfitting of obtained rule sets.