The problem of choosing the kernel for one-class support vector machines

Abstract

The article presents a review of one-class support vector machine (1-SVM) used when there is not enough data for abnormal technological object’s behavior detection. Investigated are three procedures of the SVM’s kernel parameter evaluation. Two of them are known in literature as the cross validation method and the maximum dispersion method, and the third one is an author-suggested modification of the maximum dispersion method, minimizing the kernel matrix’s entropy. It is shown that for classification without counting training data set ejections the suggested procedure provides the classification’s quality equal to the first one, and with less value of the kernel parameter.