Optimization Approaches for Parameters of SVM

Abstract

The method of SVM parameter optimization is discussed. The difference of parameter selection has an important influence on the classification accuracy of the sample. In practical systems it is difficult to obtain thousands of samples. In most cases, it can only rely on hundreds of samples to analysis and forecast. And studies have confirmed that because of the unique kernel function and classification of SVM, SVM has a greater advantage in solving small sample, nonlinear and high-dimensional pattern. So, this paper uses SVM to solve small sample classification problem. Moreover, when the parameters of SVM are optimized, higher classification accuracy can be obtained. The grid search and GA are applied to two data sets with different feature numbers, and the prediction effect is analyzed. The results show that the fewer the number of features, the better the effect of the grid search method, the more the number of features, the more obvious the advantage of GA. So GA optimizes SVM is better when higher accuracy and shorter time is required.

Jinxiang Chen and Yilan Yin are equally contributed to this work.