Experiments on Kernel Tree Support Vector Machines for Text Categorization

Abstract

Text categorization is one of the most interesting topic, due to the extremely increase of digital documents. The Support Vector Machine algorithm (SVM) is one of the most effective technique for solving this problem. However, SVM requires the user to choose the kernel function and parameters of the function, which directly effect to the performance of the classifiers. This paper proposes a novel method, named Kernel Tree SVM, which represents the multiple kernel function with a tree structure. The functions are selected and formed by using genetic programming (GP). Moreover, the gradient descent method is used to perform fine tune on parameter values in each tree. The method is benchmarked on WebKB and 20Newsgroup datasets. The results prove that the method can find a bettr optimal solution than the SVM tuned with the gradient method.