Feature Reduction for High-Precision Text Classifi- cation

Processing high dimensional features is the key of documents analysis and text classification. Traditional technologies for selecting or extracting rely heavily on the distribution of term features in the set of documents. It generally needs high computation cost to find the significant features. In this paper, we propose a new feature reduction method based on the analysis of discriminant coefficient for text classification. The main term features called the log-scaled discriminant coefficients are designed and generated efficiently. The original feature dimension of term frequency is then converted into a small size feature set. The reduced document features are used to learn classifiers using support vector machine. The experimental results show that the proposed feature reduction approach is effectiveness in comparison with the other methods. Especially, there is an obviously great improvement in precision of text classification.