Multi-view User Preference Learning with Knowledge Graph for Recommendation

To learn more comprehensive user preference, existing works in recommendation propose to utilize side information, like Knowledge Graph (KG). However, the user representation can come from many views, including ID attributes of the user, collaborative signals of interaction history, and fine-grained preferences in KG, which has not been well studied in previous works. To address the limitation, in this work, we propose the Multi-view User Preference Learning with knowledge graph (MUPL) for recommendation to address the limitation. Specifically, we propose to employ Gate Recurrent Unit (GRU) to learn the user latent collaborative feature from interaction sequence. Besides, we design a Knowledge Graph Attention Network (KGANet) to capture user fine-grained preference for the entities related with the items. Then we fusion user ID attributes, the collaborative feature and fine-grained preference for entities into the user final representation. Similarly, we employ an item encoder to get the item final representation. Finally, a predictor is proposed for recommendation. Extensive experiments on three public datasets show that our model outperforms the state-of-the-art (SOTA) methods on effectiveness.