Incremental proximal support vector classifier for multi-class classification

Abstract

Proximal support vector machine is a variation of standard support vector machine and can be trained extremely efficiently for binary classification. However in many application fields, multi-class classification and incremental learning must be supported. Incremental linear proximal support vector classifier for multi-class classification has been developed in recent years, but only its performance in one-against-all manner has been investigated, and the application of proximal support vector machine for nonlinear multi-class classification has not been studied. In order to apply proximal support vector machine to more fields, three multi-class classification policies (one-against-all, one-against-one, DAGSVM) applied to incremental linear proximal support vector classifier are compared and incremental nonlinear proximal support vector classifier for multi-class classification based on Gaussian kernel is investigated in the paper. The experiments indicate that one-against-all policy is best for incremental linear proximal support vector classifier according to the tradeoff between computing complexity and correctness, and the introduced incremental nonlinear proximal support vector classifier is effective in one-against-all manner when the reduce rate is below 0.6.