Text feature selection for relevance discovery: A fusion-based approach

This thesis presents innovative and effective feature selection models and frameworks to select and weight relevant features that describe user information needs. The proposed techniques fuse different text features to overcome problems in latent Dirichlet allocation and the relevant features discovered by existing relevance discovery algorithms. The proposed models and frameworks extend multiple random sets to model and understand the complex relationships between different entities that affect the weighting process of topical terms at both document- and collection-levels. The proposed techniques can reduce uncertainties in discovered relevant features and significantly improve the performance of text mining applications.