An Improved Feature Selection Method Based On Information Gain

Feature dimension is the basic technology in automatic text classification. And feature selection is one of the main patterns of feature dimension. This paper presents an improved information gain (IG) method based on the combination. The process of this method includes: step one, low frequency words of little importance are removed by document frequency (DF) to reduce the dependence of IG on them. Step two, the degree of representativeness to a category and the degree of differentiation among categories are proposed to improve the traditional IG method. The proposed method, compared with the traditional IG method by experiments, is proved to have a better classification effect.