Reconsideration and Experimental Study on Rule Ensemble Learning Using Kernel Method

Abstract

This study aims to learn rule ensembles, which are classifiers given by a weighted sum of if-then rules. Since the number of rules increases exponentially with respect to the size of a data set, conventional rule learning methods requires selection of useful rules. However, by using the kernel method, instead of sacrificing explicit representation of rule classifiers, we can achieve a rule ensemble considering all the rules with realistic computational complexity. In this presentation, we reconsider our previously proposed kernel-based methods for rule ensemble learning, and examine their properties through numerical experiments.