The Selection And Improvement Of Support Vector Machine Kernels

With the arrival of information era and generation of mass data,data mining technology has been developed rapidly.In order to mine useful information and hidden rules from a large amount of data efficiently,Vaplik et al.proposed support vector machine algorithm based on structural risk minimization and statistical learning theory which shows great generalization ability.Kernel selection is the focus and difficulty of SVM learning and research,and had not formed a complete system yet.The research on kernel function selection mainly focuses on two aspects:the selection of kernel function types and the selection of kernel function parameters.Choosing different type of kernel function and its parameters will make different nonlinear transformations and feature space,then different SVM training results and classification results will also be generated.This paper mainly studies the properties and classification performance of kernel functions from the following aspects:Firstly,studying the properties of four different kernel functions which was in common use respectively:linear kernel,polynomial kernel,Gaussian radial basis kernel and sigmoid kernel.Support vector machine models composed of different kernel functions within the same range of parameters are used to classify different data sets,then we will get respective classification performance.Secondary,defining two new kernel functions that satisfied the Mercer condition:inverse multiple quadratic kernel and k type kernel.The different characteristics of global kernel function and local kernel function are expounded.Based on this criterion,all kernel functions in this paper are divided into two classes.Thirdly,introducing the basic form and theoretical superiority of mixed kernel function,and a new SVM kernel function is obtained by mixing global kernel function with local kernel function using the idea of mixed kernel function.Numerical experiments show that the new mixed kernel function has better classification performance than single kernel functions.To further promote the mixed kernel function,we obtain the extended mixed kernel function and much more improved classification performance.