Feature Selection Combined Category Concentration Degree with Minimal Set Covering

摘要

Feature selection is the core research topic in text categorization. Selected feature subset directly influences results of text categorization. Firstly, word frequency and document frequency were analyzed. And then, the category concentration degree based on word frequency and document frequency was proposed. Next, set covering was introduced into rough sets and an attribute reduction algorithm based on minimal set covering was provided. Finally, a new feature selection method combined the proposed category concentration degree with the provided attribute reduction algorithm was presented. The presented feature selection method firstly uses the proposed category concentration degree to select features and filter out some terms to reduce the sparsity of feature spaces, and then employs the provided attribute reduction algorithm to eliminate redundancy, so that the more representative feature subset was acquired. The experimental results show that presented feature selection method is better than the three classical feature selection methods: information gain (IG), x~2 statistics (CHI), mutual information (MI) in time performance, macro-average F_1 and micro-average F_1.