Twin Support Vector Machines via Fast Generalized Newton Refinement

Twin SVM (TWSVM), as a computationally effective classification tool, is shown to be better than GEPSVM and SVM in favor of classification effectiveness. However, two dual QPPs arising from TWSVM leads to the higher computational time compared to GEPSVM and one has to look for approximate solutions when the data points are very large. In this paper, by slightly reformulating the primal problem of TWSVM, a new and original optimization modeling is constructed. As opposed to the TWSVM classifier, our method obtains the solution directly from solving primal problems of TWSVM using fast generalized Newton refinement method. In addition to keeping the original idea in TWSVM, still the edges of our method lie in considerably less computing time with respect to TWSVM, which is comparable to that of GEPSVM. Experiments tried out on standard datasets disclose the effectiveness of our method. Keywords: TWSVM; dual QPPs; approximate.