Research On CVM With Extensive Kernel Methods And Its Parameter C's Selection

Support Vector Machine (SVM) has became a popular machine learning method in recent years. Learning the small samples is done at the base of structural risk minimization. However the problems of algorithm speed, support vector stock and parameter selection must be addressed in training the large samples with SVM. The requirements of high speed, low memory and less support vectors are achieved difficultly in SVM, because of excessive support vectors in training. Core vector machine (CVM) is an important improvement of SVM. In CVM the support vectors are reduced greatly, but there are still defects such as restrictions on the kernel, low generalization ability and so on. Therefore, the algorithms of CVM with the extensive kernel (ECVM) and parameter selection based on kernel clustering are proposed to avoid the shortcomings of current algorithms and improve the efficiency of training and classification in SVM.Firstly, the relationship between minimal enclosing ball (MEB) and center-constraint MEB is analyzed about the kernel restriction in MEB problem. A new method in updating the center and radius is raised in the algorithm of CVM with the extensive kernel. The algorithm is proved to be convergent theoretically and is analyzed in the complexity of time and space.Secondly, the selection algorithm based on the relative distance in clustering (KCRD) is put forward to avoid the lack of current algorithm. The sample points are clustered in the feather space with the cluster algorithm, and then the parameter is obtained according to the ratio of the distance between cluster centers. In this paper, the complexity is analyzed to improve generalization ability of CVM.Finally, the experiment is done in the Linux environment.CVM, simple CVM and ECVM is compared in CPU time, the size of core vector set and test accuracy. The accuracy of prediction is analyzed in the algorithm with parameter c , which is selected by the cross-validation method, structural risk method and kernel clustering and relative distance method. From the experimental results, it is showed that ECVM is able to remove restrictions on the kernel, reduce complexity and improve the generalization ability. It is also simply inferred that the performance of the algorithm is improved with the parameter selected by the kernel clustering and relative distance method.