KEP-Rec: A Knowledge Enhanced User-Item Relation Prediction Model for Personalized Recommendation

For more accurate, diversified and interpretable personalized recommendation, the joint consideration of user-item interaction information and side information in knowledge graph has become a research hotspot. Traditional models based on collaborative filtering usually have cold start and sparse problems. The existing recommendation model based on knowledge graph can enrich the representation of users and items by using graph structure information from the knowledge graph, and make it more interpretable. Although the efforts have achieved a certain performance improvement, they consider all entities in knowledge graph globally for all users, and the aggregation strategy is single. In this paper, we propose KEP-Rec, a Knowledge Enhanced User-Item Relation Prediction Model for Personalized Recommendation. For a given target user and candidate item, KEP-Rec represents the user and item with enhanced information by knowledge graph for predicting the interacted probability between them and further personalized recommendation. In detail, KEP-Rec takes into account the changes in preferences of specific users and the differences in user perception of relations. Based on the idea of collaborative filtering, KEP-Rec selects an extended entity set of the items relevant with target user and candidate item as the initial set to propagate in knowledge graph. Moreover, KEP-Rec sets an item-aware attention mechanism to consider the interaction of candidate items with different weights given by target user's historical preferences to realize the diverse representation of the user preferences. In the propagation process alone knowledge graph, the relation embedding is considered for target user to achieve personalization. Empirical results on three real datasets of music, books, and movies show that KEP-Rec significantly outperforms state-of-the-art methods.