Mainstream knowledge graph-based recommendation models rarely consider the relationship between source nodes and target nodes when fusing high-order information, leading to the introduction of too much noise information and thus affecting recommendation performance in complex network scenarios. To address this problem, this paper proposes a knowledge graph recommendation model with integrated meta-graph neighborhoods, with the goal of reducing the impact of noise information by constructing and integrating meta-graph neighborhoods, thereby improving recommendation performance. Firstly, the model obtains the initial similar sequence of the source node based on meta-graph similarity. Then, the model enhances the initial sequence using self-attention networks and linear networks, which results in a set of enhanced feature vectors that serve as the meta-graph neighborhoods of the node. Secondly, the model designs an attention mechanism based on the user's different preferences for each meta-graph to perform convolution and aggregation on the resulting meta-graph neighborhoods. Then, the model integrates the meta-graph neighborhoods into the source node to enhance the feature representation of the source node. Finally, the model uses the inner product of the enhanced vector and the user vector as the probability of user interaction with the item, which is then utilized to complete the recommendation. Experimental results on the MovieLens-20M and Last-FM datasets show that the proposed model achieves an AUC of 97.3% and 94.3%, and an F1 score of 83.1% and 75.6%, respectively. The Recall@k at k=50 are 35.4% and 31.7%, respectively. These performance metrics outperform models such as NGCF, KGCN, LKGR, and other models. The results demonstrate that the knowledge graph recommendation model with integrated meta-graph neighborhoods is effective in improving recommendation performance.

[2024-03-07] Accepted Paper /doc/2023.12.0610/v1