Experimental Study on the Margin and Generalization of Hyper-Sphere SVM

Hyper-sphere SVM is a promising method for outlier detection, novelty discovery and so on. The margin between two kinds of samples affects the classifierpsilas generalization performance to some extent. In this paper, a generalized hyper-sphere SVM (GHSSVM) is provided and the generalization performance is studied. By introducing two parameter n and b (n>b), the unique margin which is greater than zero may be obtained and the relationship between margin and generalization is investigated by the experiments. The experimental results show the proposed classifier may have better generalization performance.