Improving recommender system via knowledge graph based exploring user preference

摘要

Knowledge graph(KG) has proven to improve recommendation performance. However, most efforts explore inter-entity relatedness by mining multi-hop relations on KG, thus failing to efficiently exploit these relations for enhanced user preference. To address this, we propose an end-to-end framework to improve the recommender system via a knowledge graph based on fusing entity relation(KGFER), which can sufficiently capture the users’ preferences. The model samples from the 1-hop neighbors and relations in KG for the item with which the user interacts, and feeds them into TransR layer. Following this, CNN is employed to learn item features from entity-relations, and then aggregate item features with the interacting item by MLP. Finally, we apply user preference to project the refined item embeddings to the user latent space to predict the potential probability of the target item in which the user is interested in. Extensive experiments on four datasets about book, movie, music, and yelp2018 demonstrate that our approach outperforms state-of-the-art baselines. Also, further experiments show that the user preference matrix indeed makes a great contribution to our approach.