A twin hyper-sphere multi-class classification support vector machine

Abstract

For the binary classification problem, twin hyper-sphere support vector machine (THSVM) is an improvement on the twin support vector machine (TSVM), it generates two hyper-spheres instead of a pair of hyper-planes. However, multi-class classification problems are always met in our real life. In this paper, we propose a new algorithm for the multi-classification problem, which is called twin hyper-sphere multi-classification support vector machine (THKSVM). Different from the “1-versus-rest” TSVM, THKSVM employs the “rest-versus-1” structure. Only one class of samples are restricted in the constraint, then it improves the computational speed of the classifier, especially for the case of the number of classes being large. Moreover, THKSVM avoids the inverse operation of a matrix when solving its dual quadratic programming problems (QPPs) compared with Twin-KSVC, so it is available for the large-scale data. Experimental results on six benchmark datasets indicate that THKSVM yields the comparable prediction accuracy with other algorithms, but it costs the shortest computational time among four algorithms.