Training One-class Support Vector Machines in the Primal Space

Abstract

In this paper, we proposed a novel Newton-type solver for one-class support vector machines in the primal space directly. Firstly, utilizing reproducing property of kernel and Huber regression function, original constrained quadratic programming is transformed into approximate unconstrained one, which is continuous and twice differentiable. Then, we give a Newton-type training algorithm to solve it. Further analysis shows the computation complexity of our algorithm is identical with theoretical lower bound for solving one-class support vector machines. In the end, experiments on 9 UCI datasets are done to validate the affectivity of proposed algorithm, and when comparing with dual method (LIBSVM),its produces comparative testing accuracy, better training speed, and less support vectors.