Biterm Tensor Topic Model for Short Reviews in Recommender System

This paper proposes the Biterm Tensor Topic Model (BTTM) for mining short review in rating prediction problems. The new method tightly integrates the Tensor Topic Model and the Biterm Topic Model. Our contributions include deriving the Maximum A Posteriori (MAP) estimates for topic probabilities tensor across different users and items, taking into account a bilinear low-rank structure for the “user-item” space and word co-occurrence patterns within the subspace. The Biterm Tensor Topic Probabilistic Matrix Factorization (BTTPMF) is also introduced for joint modeling of ratings and reviews. Experimental results demonstrate that BTTM effectively uncovers potential semantic topics from short reviews and predicts ratings more accurately than baseline models.