On Canonical Modelling Alternatives for Probabilistic Multi-Class Support Vector Machines

Support vector machines are a flexible tool for binary classification. Two different paradigms have been proposed to extend their use in multi-class classification. Firstly, one can formulate a single quadratic problem that can be used for multi-class classification. Secondly, one may reduce the problem to a set of binary problems and use probabilistic modelling to fuse results from multiple classifiers. Both approaches entail making non-canonical choices. In our paper we empirically evaluate whether alternative methods in probabilistic modelling yield better classification results. We focus on three factors - the prior used in Platt's modelling method, the coupling method used to combine probabilistic predictions from models fitted to decision values of binary support vector machine models, and the choice of the scoring function. We show that by employing alternative approaches one may gain additional increase in the accuracy.