Research On Robust One-Class Support Vector Machines

One-class classification is a common task in the fields of machine learning and pattern recognition.It can effectively solve the problems of only one-class samples in the given training samples and extreme class imbalance.At present,many scholars have proposed lots of methods for solving one-class classification problems.The two commonly used methods are one-class support vector machine(OCSVM)and support vector data description(SVDD).However,OCSVM has some defects,e.g.sensitive to outliers existing in the training set,and poor robustness.Based on this,this thesis studied from two aspects,i.e.,sample weighting and improved loss function.1.Adaptive weighted one-class support vector machine(AWOCSVM)is proposed.The adaptive weights of the proposed method are based on the distances between the sample points and the sample center together with the local density of the samples.Therefore,the outliers may get smaller weights,which weakens their influences on the decision boundary and improves the robustness of OCSVM.The feasibility of the proposed method is verified on the synthetic and benchmark data sets.2.Robust one-class support vector machine with the rescaled hinge loss function(ROCSVM-RHHQ)is proposed.The proposed method uses non-convex and bounded loss function to instead unbounded hinge function,which effectively reduces the negative effects of the outliers and improves the robustness of OCSVM.Then,performance evaluation and robustness analysis are carried out theoretically.In addition,the comparison between ROCSVM-RHHQ and the other three related approaches are conducted on the synthetic and benchmark data sets.Experimental results show that the proposed method has better robustness.