Research On Robust Least Square One-class Support Vector Machines

One-class support vector machine is regarded as a commonly used one-class classifier.In comparison with the conventional one-class support vector machine,least square one-class support vector machine(LS-OCSVM)can more accurately describe the similarity between a new-coming sample and the given training set.However,LS-OCSVM is very sensitive to the outliers in training set.The main reason lies that the values of squared error function for these outliers are larger.Hence,LS-OCSVM prefers these outliers.To enhance the classification performance of LS-OCSVM,researches are carried out upon LS-OCSVM in this thesis.Two variants of LS-OCSVM are proposed.1.Robust least square one-class support vector machine with capped L1-norm(RLSOCSVM-CL)is proposed.In the proposed RLSOCSVM-CL,the loss for each sample is upper bounded.Thus,the larger losses of the outliers in training set are constrained within a smaller range to alleviate the negative effect of these outliers upon hyperplane.Therefore,the robustness of LS-OCSVM against outliers can be improved.To validate the efficiency of the proposed method,the comparisons between it with its related approaches upon the synthetic and UCI benchmark data sets are conducted.Experimental results demonstrate that RLSOCSVM-CL possesses the better anti-outlier ability.2.Robust least square one-class support vector machine with C-loss function,(RLS-OCSVM)is proposed.RLS-OCSVM uses nonconvex and bounded C-loss function to substitute the unbounded squared loss function of LS-OCSVM to effectively suppress the loss caused by outliers.Hence,the robustness of RLS-OCSVM against outliers can be enhanced.To verify the efficiency of RLS-OCSVM,the comparisons between it with its three related methods on the synthetic,UCI benchmark,and handwritten data sets are conducted.Experimental results show that the proposed RLS-OCSVM is more robust against outliers.