MK-MSMCR: An efficient multiple kernel approach to multi-class classification

Abstract

This paper introduces a novel multi-class support vector classification and regression (MSVCR) algorithm with multiple kernel learning (MK-MSVCR). We present a new MK-MSVCR algorithm based on two-stage learning (MK-MSVCRTSL). The two-stage learning aims to make classification algorithms better when dealing with complex data by using the first stage of learning to generate ?representative? or ?important? samples. We first establish the fast learning rate of MKMSVCR algorithm for multi-class classification with independent and identically distributed (i.i.d.) samples amd uniformly ergodic Markov chain (u.e.M.c.) smaples, and prove that MK-MSVCR algorithm is consistent. We show the numerical investigation on the learning performance of MK-MSVCR-TSL algorithm. The experimental studies indicate that the proposed MK-MSVCR-TSL algorithm has better learning performance in terms of prediction accuracy, sampling and training total time than other multi-class classification algorithms.