Successive Overrelaxation for Twin Parametric-margin Support Vector Machines

摘要

The Twin Parametric-margin Support Vector Machine (TPMSVM) is an effective kernel-based tool for binary classification. It constructs a pair of non-parallel parametric-margin hyperplanes by solving two Quadratic Programming Problems (QPPs), which lead higher computational costs. In this paper, we propose a novel improved version of TPMSVM, called ITPMSVM. Our ITPMSVM is inspired by TBSVM, and the main advantages of our ITPMSVM are: (1) In primal problems of our ITPMSVM, the additional regularization terms b_+~2 and b_-~2 are added to guarantee the positive definition of the QPPs. (2) By introducing the new additional terms, it leads to the unique solution in our ITPMSVM, where is not unique in TPMSVM. (3) Since our dual problems only have the boundary conditions, an effective technique (successive overrelaxation, SOR) is applied to our ITPMSVM. The experimental results on a series of benchmark datasets confirm the merits of our ITPMSVM.