Standard measure and SVM measure for feature selection and their performance effect for text classification

This paper compares the prediction performance of document classification based on a variety of feature selection measures. Empirical experiments were conducted for the dataset re0 with 10 measures for feature selection and with SVM. It is confirmed that the feature selection based on the SVM-score proposed by Sakai and Hirokawa (2012) outperforms the standard measures with small number of features. In fact, 100 words are enough to get the similar performance obtained with all words. The reason of good performance of this feature selection is that the SVM-score capture not only the characteristic words of positive samples but of negative samples as well.