Enhanced Implicit Collaborative Knowledge Graph for Recommendation

Abstract

Knowledge-aware Recommendation (KGR) with Graph Neural Network (GNN) aggregates information on the knowledge graph to capture high-order knowledge representations for the recommendation task, which becomes the mainstream approaches in the field of KGR. While these studies have achieved great success, we emphasize that strict supervision signals of training can have negative impacts on the model. Specifically, the supervision signals force the model to align constantly with sparse positive interaction data, which can result in the loss of valuable knowledge semantics. Moreover, it may also weaken model's ability to explore the potential interests for user. To address the aforementioned issues, we propose a new model named Enhanced Implicit Collaborative Knowledge Graph for Recommendation (EICKR), which uses a multi-task learning schema to extract user interests. This model calculates implicit relations among users, items, and entities to structure rich implicit features across different views. Subsequently, it leverages the knowledge graph and users' historical interaction data to generate an implicit collaborative knowledge graph, bridging the semantic difference between the knowledge graph and interaction data. Furthermore, a graph-enhanced completion task based on implicit relations is introduced to explore unobserved yet valuable interaction features among users, items, and entities, which can be complementary to the recommendation task. We conduct extensive experiments on benchmark datasets, which demonstrates the effectiveness of our approach and its components.