Dual collaborative topic modeling from implicit feedbacks
Gai Li, Lei Li · 2014
The research on implicit feedback collaborative filtering is also called One-Class Collaborative Filtering (OCCF). In this paper, we propose a Dual Collaborative Topic Regression (DCTR) model to integrate an implicit feedback rating matrix (using probabilistic matrix factorization) into the user's/item's content information (using a topic model) to improve OCCF accuracy. The matrix factorization of the implicit feedback rating matrix learns the low-rank user's/item's latent feature space, while the topic modeling provides a content representation of the user/item in the user's/item's latent topic space. Then we use the user's and item's latent topic space to construct their latent feature space. It can be seen that the Collaborative Topic Regression (CTR) model and Probabilistic Matrix Factorization (PMF) model can be derived as special cases from DCTR. We provide a scalable, linearly complexity model fitting procedure through coordinate ascent optimization which can dramatically reduce the computation cost in every iteration. The experimental results on two datasets show that DCTR outperforms the state-of-the-art models. More importantly, our model can solve the new-user and new-item cold-start problems simultaneously.