A New Feature Selection Algorithm Based on Category Difference for Text Categorization.

The feature selection is an important step which can reduce the dimensionality and improve the performance of the classifiers in text categorization. Many popular feature selection methods do not consider the difference in the distribution of different categories on a feature. In this paper, we propose a new filter based feature selection algorithm, namely fused distance feature selection (FDFS), which evaluates the significance of a feature by taking account of the difference in the distribution of different categories and selects more discriminative features with the minimal number. The proposed algorithm is investigated both inside and outside perspectives on four benchmark document datasets, 20-Newsgroups, WebKB, CSDMC2010 and Ohsumed, using Linear Support Vector Machine (LSVM) and Multinomial Naïve Bayes (MNB) classifiers. The experimental results indicate that our proposed method provides a competitive result, where its average ranking is 1.25 on LSVM and 1 on MNB.