Relative term-frequency based feature selection for text categorization

Automatic feature selection methods such as document frequency (DF), information gain (IG), mutual information (MI) and so on are commonly applied in the preprocess of text categorization in order to reduce the originally high feature dimension to a bearable level, meanwhile reduce noise to improve precision. Generally they assess a specific term by calculating its occurrences among individual categories or in the entire corpus, where "occurring in a document" is simply defined as occurring at least once. A major drawback of this measure is that, for a single document, it might count a recurrent term the same as a rare term, while the former term is obviously more informative and should less likely be removed. In this paper we propose a possible approach to overcome this problem, which adjusts the occurrences count according to the relative term frequency, thus stressing those recurrent words in each document. While it can be applied to all feature selection methods, we implemented it on several of them and see notable improvements in the performances.