Collaborative Topic Modeling for Text Tensors

A variety of generative topic models have been successfully applied to model corpus of documents with continuous metadata. But there is no efficient model dealing with documents having a user-item-word structure. This structure forms a 3-way text tensor, and texts correlate with each other through users and items. In this paper we propose an elegant Tensor topic model (TTM) for text tensors inspired by Tucker decomposition, in which both user and item dimensions are co-reduced together with vocabulary space. So we get low-rank representations for not only words but also users and items from TTM. Also, general rules are developed to transform decomposition model into a probabilistic one. Experiments show that TTM outperforms existing topic models in modeling texts with a user-item-word structure.