Constructing Kernels for One-Class Support Vector Machine

OCSVM (one-class support vector machine) is a variant of SVM which only use positive class sample set in training. Since only positive samples can be used in OCSVM, Fully exploiting and using the features of the training samples is of great significance to improve its classification performance. Thus, two aspects of study on kernels have been done in this paper: first, we propose a kernel constructing method called WFCD (weighted feature-contribution-degree) kernel constructing method, in which a PCA (principal component analysis) is performed to the training samples to obtain a vector set with the dimension being sorted by corresponding Eigen values and then using this vector set to apply a weighed kernel method to concentrate on the larger Eigen value dimensions, second, we employ the Fisher kernel in OCSVM to decide whether a kernel constructed based on the training sample set has better performance. Experimental results on UCI standard data sets indicate that our method outperforms the general kernel methods and promotes the classification effect considerably.