Sparse Learning for Linear Twin Parameter-margin Support Vector Machine

Twin Parameter-margin support vector machine (TPMSVM) is a recent very powerful binary classifier. To improve its sparsity, a linear sparse TPMSVM (Lin-STPMSVM) is proposed in this paper. In the primal problem, the vectors defining the hyperplane are replaced with their expression in terms of the dual variables as derived from Karush Khun Tucker (KKT) conditions. Then the new primal problems are directly optimized, thus ensuring the sparsity of the solutions. Numerical experiments show that the solution obtained by new model is more sparse without reducing the accuracy. Therefore, Lin-STPMSVM not only inherits the advantages of TPMSVM, but also has the characteristics of sparsity, stability and robustness in dealing with classification problems.