Probabilistic large margin machine

Large margin learning has been widely applied in solving supervised classification problems. One representative model in large margin learning is the support vector machine (SVM). As the linear classification constraints in the SVM optimization problem are determined with certainty, the performance of SVM is limited. In this study, we propose a new large margin learning model, named probabilistic large margin machine (PLMM), with the linear classification constraints bounded by probabilistic thresholds. In comparison with the SVM, the PLMM incorporates the prior probabilities and the distribution information of each class into the decision hyperplane learning. Mathematically the optimization problem involved in the PLMM can be treated as only one second order cone programming (SOCP) problem, which can be solved efficiently. The experimental results demonstrate the effectiveness of the PLMM model.