A Novel Classification-Rejection Sphere SVMs for Multi-class Classification Problems

Abstract

A new algorithm called classification-rejection sphere support vector machines (C-R sphere SVMs) is proposed based on the human thoughts of recognition and support vector machine (SVM) technology for multi-class classification problems. The new algorithm constructs a classifying sphere for each class instead of a minimum sphere. Like human being, C-R sphere SVMs can not only classify the multi-class data but reject the data which do not belong to any class known. In comparison with hyperplane SVMs, the algorithm can construct a new classifying sphere for a new class without affecting other classifying spheres so that it can reduce computational complexity obviously. The effect of the increment coefficient lambda and Gaussian kernel parameter sigma on the performance of C-R sphere SVMs is analyzed. Numerical simulations are performed on a real dataset (from the UCI dataset repository). The results show that the C-R sphere SVM algorithm exhibits good performance when appropriate values of lambda and sigma are taken.