Categorical Term Frequency Probability Based Feature Selection for Document Categorization

Document categorization technology heavily relies on the categorical distribution of features. Those terms which occur unevenly in various categories have strong distinguishable information as to categorization. At first, we give the definition of CTFP (Categorical Term Frequency Probability), which will be used to accurately reflect the categorical characteristics of terms on each category. Then, the CTFP_VM (Variance-Mean based on CTFP) feature selection criterion is introduced to reveal the category distribution difference. After computing and ranking the variance mean based on CTFP distribution for each term, feature sets are obtained for document categorization. We perform the document categorization experiments on SVM classifiers with the well-known Reuters-21578 and 20news-18828 corpuses as unbalanced and balanced corpus respectively. Experiments compare the novel methods with other conventional feature selection algorithms and the proposed method achieves the best feature set for document categorization The experimental results also demonstrate that the proposed variance mean feature selection method base on CTFP not only has better F1-metric for document categorization but excellent corpus adaptability.