A Learning Approach for Fast Training of Support Vector Machines

In this paper, we propose a learning method for fast training of support vector machines (SVMs). First, we divide the two-class training samples into two sets according to the labels. Secondly, the two set one-class samples are trained by using one-class SVM (OCSVM) respectively, and we get two set support vectors (SVs). Finally, the two set SVs are combined into a set of two-class training samples and trained by normal SVM algorithm. The experimental results show the proposed method can improve the training speed and generate the simpler decision function, at the same time the accuracy is kept.