Text feature selection approach by means of class difference

Feature selection aims to seek some relevant features from the whole feature space to construct a feature subset and it can help us to handle clustering, classification and retrieval. This paper considers feature selection for text categorization. We put forward a filter feature selection scheme based on class difference measure. The key idea of our proposed algorithm is difference between the frequencies of document of class in which a term occurs. We compared the suggested approach with information gain and chi-square statistic that are usually utilized in text feature selection. The experimental results demonstrated that the new criterion was simpler and produced better performance in most of the cases.