MI-KGNN: Exploring Multi-dimension Interactions for Recommendation Based on Knowledge Graph Neural Networks

To achieve more accurate recommendations, a consensus of the research community is that not only explicit information (i.e., historical user-item interactions) but also implicit information (i.e., side information) should be utilized. Generally, both explicit and implicit information can be categorized according to the following assumptions: 1) Users with same behaviors are similar; 2) Items related to the same user are similar; 3) Items with same attributes are similar; and 4) Users with same interests are similar. However, none of existing studies has fully explored such information. To this end, we put forward Multi-dimension Interactions based Knowledge Graph Neural Networks (MI-KGNN), i.e., a GNN-based recommendation model that characterizes the similarity between users and items through embedding propagation in the knowledge graph. Specifically, apart from the traditional user-item and item-user interactions, we define another two types of interactions by introducing three different bipartite graphs. On one hand, we explore the interaction between items and the neighborhood during the information aggregation process. On the other hand, we explore the interaction between users and the neighborhood during embedding propagation. These interactions allow information to propagate in the direction indicated by the above four assumptions. In such a way, MI-KGNN effectively extracts both semantic information and structural information in the knowledge graph. Experimental results show that MI-KGNN significantly outperforms state-of-the-art methods in top-K recommendations.