A Study Of Kernel Classification Algorithm Based On Double-Kernel Combinition

With the successful application of Support Vector Machine (SVM), kernel methods, as an important component of SVM, have caused more and more researchers'attention nowadays, and are widely used in pattern recognition, image processing, machine learning and many other fields. Along with the further development of kernel methods, researchers realized that traditional kernelised classification methods are limited in their performance on complicated data sets because of the using of a single and fixed kernel.In order to overcome the limitation of traditional kernelised classification methods, a novel optimal double-kernel combination method (ODKC) is proposed in this paper. ODKC is composed of two different basic kernels, which can take the advantages of various kernels to better adapt to complicated tasks.The work in this paper is as follow:(1). A kind of criterion function is proposed to seek the optimal double-kernel combination. The problem of optimization with this criterion function can be reduced to a generalized eigenvalue problem easily, so that ODKC processes non-iterative and low computational complex property.(2). A unified framework is proposed, in which we study three kinds of combinations of mappings. Firstly, data sets are mapped by two basic kernels into different feature spaces respectively, and then three kinds of optimal composite kernels are constructed by integrating information of the two feature spaces. It is the unified framework that we can discuss and compare the effect of each kind of combination easily.(3). Two parts of experiments on five data sets are conducted to demonstrate the effectiveness of our methods.