Research On Locality Weighting One-Class Support Vector Machines

One-class classification is regarded as a machine learning task between supervised learning and unsupervised learning.It can effectively solve the problem of training with samples in only one class and the problem of extreme class imbalance.Till now,a large number of one-class classification methods have been proposed.O ne of the commonly used methods is one-class support vector machine(OCSVM).However,the traditional OCSVM has some shortcomings.For example,OCSVM has not considered the influence of the geometric distribution of the given training samples upon the classification performance.Therefore,in this thesis,the locality geometric information of the given training samples are utilized to conduct the research from two aspects,i.e.,margin improvement and weighting misclassified samples for OCSVM.1.Locality correlation preserving based one-class support vector machine(LCP-OCSVM)is proposed.The proposed method introduces locality correlation preserving(LCP)into the traditional one-class support vector machine,which inherits the merits of LCP and OCSVM.Therefore,LCP-OCSVM can keep locality correlation of the normal data and margin maximization between the normal data and the origin in the high-dimensional feature space.Experimental results on the synthetic and benchmark datasets verify the practicability of the proposed method.2.Locality preserving weighed one-class vector machine(LPWOCSVM)is proposed.In order to reduce the influence of misclassified samples on the classification boundary o f OCSVM,the proposed method construct locality preserving weighed vector by utilizing the locality geometric information of the given training samples.The misclassified samples are assigned with smaller weights by the above-mentioned locality preserving weighting vector.Therefore,the classification boundary of OCSVM may become more compact.Experimental result on the synthetic and benchmark datasets demonstrate that the proposed method is more robust against noise and has better generalization performance.