Feature selection for support vector machines in text categorization

Support vector machine (SVM) has been successfully applied to text categorization. We find that the decision hyperplane of SVM classifier can also be used for feature selection. When linear kernel is used, the decision rule can be seen as a combination of weighted values of each dimension of the vector to be categorized, each weight indicating the importance of corresponding dimension in the categorization decision. Hence we can only preserve those dimensions with bigger weights. In this way we combine the feature selection into the learning process of SVM training. Empirical experiment result is given to verify our new method.