A Cross-Domain Latent Topic Model for Item Tagging and Recommendation Systems

Cross-domain data analysis is becoming increasingly crucial in media convergence. Most of the existing methods implement cross-domain feature unifying by numerically fitting the experimental results, which lack interpretability. To this end, we propose a cross-domain latent topic model (CDLT) via Latent Dirichlet Allocation and evaluate our model on two applications. First, the CDLT model exploits the supervised tag information to learn the unified topic space while obtain domain discrimination. Second, we put forward a cross-domain item tagging method (CDLT-Tag) based on CDLT model. Third, combining user behaviors and CDLT model, we propose a cross-domain recommendation algorithm (CDLT-Rec). Meta-learning is used to estimate the user preference on new unseen domains that alleviate the cold-start problem. This paper proves the effectiveness of CDLT model by comparing these two applications with existing tagging and recommendation algorithms in a cross-domain scenario on both Movielens and Douban dataset.