Hybrid Recommender Systems with Deep Learning

As one of the most popular recommendation algorithms, collaborative filtering (CF) suggests items favored by like-minded based on user ratings. However, CF performs worse for users and items with fewer ratings, which is known as the cold-start problem. On the other hand, the auxiliary information of items such as images and reviews can be helpful for relieving the cold-start issue and improving recommendation accuracy. How to effectively extract features from heterogeneous auxiliary information and integrate them with collaborative filtering remains a big challenge. In this thesis, we propose a tightly-coupled hybrid recommender system named Fusion-MF-Mix via a deep fusion framework, which extracts features automatically from different domains and enables two-way information propagation between feature learning and rating prediction. We conduct extensive experiments to evaluate our model on two large-scale real-world datasets from amazon movie and book recommendation domains. The results show that our model outperforms other baseline models