Study on Multi-class Text Classification Based on Improved SVM

Traditional SVM multi-class classification methods such as 1-a-r, 1-a-1, and DAG-SVM are popular learning techniques, but they often have the problems that the input text instances are nonlinear separable or their training/testing process is time consuming. In this chapter, we propose an improved SVM multi-class classification algorithm: first, the nonlinear separable texts in input space are mapped into a high dimensional feature space (Hilbert space) by using Mercer kernel function for obtaining linear separable texts; then in the Hilbert space, we construct SVM multi-class classifiers with binary tree to recognize testing instances. Our experiments demonstrate that for some web text classification problems, the proposed method can effectively solve the nonlinear separable problem existing in input text space, saves training/testing time efficiently, and improves the text classification accuracy.