Feature Selection Method Based on Feature Distinguishability and Fractal Dimension

Abstract

Feature selection is the core step in text categorization system. The selected feature subset directly influences the results of text categorization system. It firstly analyzed simply several classic feature selection methods and summarized their deficiencies, and then presented feature distinguishability. Subsequently, it introduced fractal dimension into rough sets and provided an attribute reduction algorithm based on fractal dimension. Finally, it combined the reduction algorithm with the feature distinguishability and proposed a comprehensive feature selection method. The comprehensive method firstly uses the feature distinguishability to select features and filter out some terms to reduce the sparsity of feature spaces, and then uses the proposed attribute reduction algorithm to eliminate redundancy, so that it can acquire the feature subset which are more representative. The experimental results show that the comprehensive method is promising.