Topic model-based recommender systems and their applications to cold-start problems

Recommender systems provide information and items that match a user's preference. This study proposes hybrid recommender models that use content-based filtering and latent Dirichlet allocation (LDA)-based models. The proposed models are extensions of the LDA where the words correspond to user characteristics and item features and are found to be suitable for handling cold-start problems, as it provides predicted ratings for new users and items via its latent dimension. These models have the advantage of analyzing item topics, item feature topics, and user characteristic topics simultaneously. Experiments conducted with the MovieLens 1M dataset illustrate that the proposed models provide similar prediction performances as baseline recommender models and are superior to the baseline models regarding the interpretability of the user and item topics.

Automated summary

This study proposes hybrid recommender models that use content-based filtering and latent Dirichlet allocation (LDA)-based models to address the cold-start problem in recommender systems. Experimental results demonstrate that these models achieve similar prediction performances compared to baseline models, while providing better interpretability of user and item topics.