Feature Selection with Structural Sparse Mode for Text Categorization

The grouped structure has successfully been embedded in sparse models for feature selection; however, some groups generated by clustering method might be difficult to interpret their semantic information if the number of words in the group is very large. This paper proposes a novel approach in which a group structure is constructed and its corresponding sparse model is used to select features for text categorization. After variable preselection, an algorithm is developed to generate groups, such that each group only contains two or three closely related words, which can reflect more essential semantic meaningful. Finally, structural sparse mode is used to select feature in wrapper way. The experimental results demonstrate that the proposed method achieves comparable precision and improves the sparsity considerably, which means that the model has better interpretability.