A new fast training algorithm for SVM

A fast SVM training algorithm is proposed in this paper. By integrating kernel caching, shrinking and using second order information, a fast Quadric Programming(QP) trainer is achieved. For traditional two-class SVM, the generalized error bound derived from Statistical Learning Theory(SLT) is computed and minimized for the selection of parameters, with the Zoutendijk(ZQP) idea and parallel method to speed up the process. For one-class SVM, a compression criterion is proposed to search the best kernel width automatically. Experiments demonstrate that the proposed method is significantly faster than LibSVM and requires less support vectors to achieve good classification accuracy.