An empirical evaluation of linear and nonlinear kernels for text classification using Support Vector Machines

This paper compares the performance of linear and nonlinear kernels of Support Vector Machines (SVM) used for text classification. The study is motivated by the previous viewpoint that linear SVM performs better than nonlinear one, and that, although there are many investigations have proved that SVM performs well in text classification, there is no serious investigation on the comparison between linear SVM and nonlinear SVM. In our study, we carry out two experiments with different datasets and use grid-search on the selection of kernel parameters. Empirical results show that, in fact, nonlinear SVM performs better than linear SVM as long as with appropriate kernel parameters. This conclusion will provide useful guidance for people applying SVM to text classification and other corresponding fields.