Multi-objective search for comprehensible rule ensembles

Abstract

We present a methodology for constructing an ensemble of rule base classifiers characterized not only by a good accuracy of classification but also by a good quality of knowledge representation. The base classifiers forming the ensemble are composed of minimal sets of rules that cover training objects, while being relevant for their high support, low anti-support and high Bayesian confirmation measure. The population of base classifiers is evolving in course of a bi-objective optimization procedure that involves accuracy of classification and diversity of base classifiers. The final population constitutes an ensemble classifier enjoying some desirable properties, as shown in a computational experiment.