Abstract: To achieve feature reduction in text categorization, the scatter difference criterion is improved to satisfy a broad range of text categorization problems using kernel commutation in the pre-treatment. A kernel-based nonlinear method is proposed to extract features. By kernel commutation, the stylebook categorization problem is solved with less linear separability. Dimension of the feature space is significantly reduced without incurring excessive information loss. Experiments show that performance of the proposed method is better than maximal scatter difference with an efficiency improvement of 4.7% for the value of F1.

Key words: text categorization, feature extraction, scatter difference, kernel commutation