Feature Selection for the Topic-Based Mixture Model in Factored Classification
Qiong Chen · 2006
Topic-based mixture model (TBMM) is a learning algorithm for factored classification. In factored classification, the class label is factored into a vector of class features. For example, the class label for a personal Web page at a university might be described by two features: the academic discipline of the person, and their position (e.g., 'chemistry professor' or 'physics student'). An approach to factored classification of text documents in which each document is assumed to be generated by a mixture of class features was proposed. Experiments in factored text classification problems show TBMM can outperform other two approaches for categories with especially sparse training data. In this paper, we analyze the feature selection for TBMM. For TBMM the feature space can be reduced to small number of feature terms with a significant improvement to classification accuracy. We present empirical results that indicate that TBMM is an adequate method to determine the feature terms for the supervised classification task