Multi-class pattern classification based on a probabilistic model of combining binary classifiers

Abstract

We propose a novel probabilistic model for constructing a multi-class pattern classifier by weighted aggregation of general binary classifiers including one-versus-the-rest, one-versus-one, and others. Our model has a latent variable that represents class membership probabilities, and it is estimated by fitting it to probability estimate outputs of binary classfiers. We apply our method to classification problems of synthetic datasets and a real world dataset of gene expression profiles. We show that our method achieves comparable performance to conventional voting heuristics. © Springer-Verlag Berlin Heidelberg 2005.