TrueSkill-based pairwise coupling for multi-class classification

Abstract

A multi-class classification problem can be solved efficiently via decomposition of the problem into multiple binary classification problems. As a way of such decomposition, we propose a novel pairwise coupling method based on the TrueSkill ranking system. Instead of aggregating all pairwise binary classification results for the final decision, the proposed method keeps track of the ranks of the classes during the successive binary classification procedure. Especially, selection of a binary classifier at a certain step is done in such a way that the multi-class classification decision using the binary classification results up to the step converges to the final one as quickly as possible. Thus, the number of binary classifications can be reduced, which in turn reduces the computational complexity of the whole classification system. Experimental results show that the complexity is reduced significantly for no or minor loss of classification performance. © 2012 Springer-Verlag.