Margin-based Diversity Measures for Ensemble Classifiers

Abstract

The classifier ensembles have been used successfully in many applications. Their superiority over single classifiers depends on the diversity of the classifiers forming the ensemble. Till now, most of the ensemble diversity measures were derived basing on the binary classification information. In this paper we propose a new group of methods, which use the margins of individual classifiers from the ensemble. These methods process the margins with a bipolar sigmoid function, as the most important information is contained in margins of low magnitude. The proposed diversity measures are evaluated for three types of ensembles of linear classifiers. The tests show that these measures are better at predicting recognition accuracy than established diversity measures, such as Q or disagreement measures, or entropy.