Evaluation of Diversity Measures for Binary Classifier Ensembles

Abstract

Diversity is an important consideration in classifier ensembles, it can be potentially expolited in order to obtain a higher classification accuracy. There is no widely accepted formal definition of diversity in classifier ensembles, thus making an objective evaluation of diversity measures difficult. We propose a set of properties and a linear program based framework for the analysis of diversity measures for ensembles of binary classifiers. Although we regard the question of what exactly defines diversity in a classifier ensemble as open, we show that the framework can be used effectively to evaluate diversity measures. We explore whether there is a useful relationship between the selected diversity measures and the ensemble accuracy. Our results cast doubt on the usefulness of diversity measures in designing a classifier ensemble, although the motivation for enforcing diversity in a classifier ensemble is justified.