Learning Topic-Based Mixture Models for Factored Classification

We present a learning algorithm for factored classification, employing a topic-based mixture model. 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’). We present an approach to factored classification of text documents in which each document is assumed to be generated by a mixture of class features. This formulation allows building on recent work on topic-based mixture models for unsupervised text analysis. We present an algorithm for supervised learning of mixture models for factored classification. Experiments in two factored text classification problems (classifying web pages and classifying the intent of email senders) demonstrate our approach, and show it can outperform earlier approaches for categories with especially sparse training data.