Research and improvement for feature selection on naive bayes text classifier

An effective feature selection is very important for an classifier. Improved feature selection method can enhance its classifier efficiency in the practical test validates. This paper studies the principle·, merits and limitations of the prevalent feature selection method. Then, the paper adopts two-stage selection modulus which is calculated by the position of paragraph and sentences respectively, and takes feature variance of two phases into consideration. Finally, the paper adopts improved algorithm in Spam Filter categorization, a quite typical text classification. Experiments show that this method works more effectively than only using mutual information method applied in Naïve bayes in selecting those representative features.