Research On Parameter Optimization And Feature Selection In Support Vector Machine Based On Gaussian Kernel With Multiple Widths

Based on statistical learning theory,SVM is a widely used machine learning classification algorithm.However,in the specific issues,kernel function and its parameters has always been the critical factor to decide generalization capability of SVM.In this paper,we introduce the GKMW which has the stronger generalization ability to replace the traditional Gauss kernel function to solve the problem that width of Gauss kernel limits the generalization scale of SVM.For the parameter optimization of GKMW,traditional gradient algorithm depend on initial value too much,and is easy to fall into local optimum.So this paper adopt GEP to optimize the problem.In order to seek parameters faster,the paper designed GEPCS algorithm--an upgraded GEP algorithm to reduce the computational cost of parameter optimization and improve the operation efficiency.Main research contents of this article are as follows:1.The parameter optimization of the GKMW is essentially a multimodal combination optimization problem,so it is more appropriate to optimize with the evolutionary computation than the traditional gradient algorithm.Based on GKMW,this paper proposed the concept of Class Separability Criterion.The GCSC criterion can determine the degree of dispersion among the categories in the corresponding characteristic space under the fixed parameters combination.And taking it as optimization direction of parameter combination.Further more,this paper introduce the criterion into the GEP algorithm,reform the original fitness function formula,and solve the problem that original fitness function will spend a lot of time training SVM.The improved GEPCS algorithm will solve the problem of the parameter optimization of GKMW,and at the same time enhance algorithm operational efficiency.2.Facing the complex data scale in the large data age,SVM often encountered data sets with redundant or unrelated features when dealing with large-scale data.In order to further improve the classification performance of SVM,this paper based on the parameters optimization of the above GKMW and the characteristics that GKMW can not only reflect the difference in the contribution of the various features to the classification,but also distinguish the characteristics of the importance of each feature in the sample,proposed a SVM feature selection algorithm on the basis of GKMW to reduce the complexity of the feature space and improve the classification performance.3.This paper applies the GEPCS algorithm to a standard UCI data sets and then conducts the classification experiment of SVM.On the basis of verifying the feasibility and effectiveness of the parameter optimization algorithm,we make a comparative experiment with traditional parameter optimization method.And the results show that the GEPCS algorithm can find optimal parameter combination,which makes the classification accuracy of the SVM approaching or exceeding the theoretical accuracy of the classified data sets.Compared with other methods,the presented method has the advantage of seeking parameters in shorter time.Then we use feature selection of data sets on the basis of parameter optimization results,the experiment shows that this feature selection algorithm can raise the classification performance of SVM,which provides a new and efficient method for parameter optimization and feature selection for SVM collocated with GKMW.