Support vector machines based composite kernel

Abstract

In order to raise the adapbility of SVM classification to the specific dataset, a composite kernel is proposed and introduced into SVM, and the parameters are optimized according to “Fisher Discriminant” and “Kernel Alignment”, to maximize the class separability in the empirical feature space and, make composite kernel to be more relevant for the dataset and adapt itself by adjusting its composed coefficient parameters, thus allowing more flexibility in the kernel choice. The performance of support vector machines based composite kernel (CK-SVM) is extensively evaluated on five UCI standard datasets, at the same time, we compare CK-SVM with other existing method and get convincing results, which reveal that the proposed method is a robust and promising classifier.