Kernels for One-Class Support Vector Machines

Abstract

One-class support vector algorithms such as OCSVM and SVDD have been successfully applied to many One-Class Classification (OCC) problems. Many authors assume that kernels like the ones used in standard binary SVM classification are also appropriate to one-class classification. However, a review of the literature indicated that in general, only the Gaussian RBF kernel gives satisfactory results in OCC problems. Nonetheless researchers are continuing unsuccessfully to try other kernel functions such as polynomial and sigmoid. In this paper, we propose to investigate whether this kernel function is the only suitable one, or whether other ones may also be appropriate for OCC. The results of our experiments on standard data-sets by using the commonly used kernels, show that the best performances are always obtained with decreasing RBF kernels such as the Gaussian kernel.