Binary classifiers ensemble based on Bregman divergence for multi-class classification

We propose a novel integration method of binary classifiers for multi-class classification. The proposed method is characterized as a minimization problem of a weighted mixture of Bregman divergence, and employs binary classifiers as class-dependent feature vector. We discuss the statistical properties of the proposed method and the relationship between the proposed method and existing multi-class classification methods, and reveal that many of the existing methods can be formulated as special cases of the proposed method. Small experiments show that the proposed method can effectively incorporate information of multiple binary classifiers into the multi-class classifier.