Jointly Learning Propagating Features on the Knowledge Graph for Movie Recommendation

Knowledge graphs are widely used as auxiliary information to improve the performance in recommender systems. This enables items to be aligned with knowledge entities and provides additional item attributes to facilitate learning interactions between users and items. However, the lack of user connections in the knowledge graph may degrade the profiling of user preferences, especially for explicit user behaviors. Furthermore, learning knowledge graph embeddings is not entirely consistent with recommendation tasks due to different objectives. To solve the aforementioned problems, we extract knowledge entities from users' explicit reviews and propose a multi-task framework to jointly learn propagating features on the knowledge graph for movie recommendations. The review-based heterogeneous graph can provide substantial information for learning user preferences. In the proposed framework, we use an attention-based multi-hop propagation mechanism to take users and movies as center nodes and extend their attributes along with the connections of the knowledge graph by recursively calculating the different contributions of their neighbors. We use two real-world datasets to show the effectiveness of our proposed model in comparison with state-of-the-art baselines. Additionally, we investigate two aspects of the proposed model in extended ablation studies.