Research on recommendation algorithm based on heterogeneous graph neural network

This paper presents the construction of a novel recommendation algorithm that leverages the principles of a heterogeneous graph neural network (HGNN). Through the in-depth study of heterogeneous graph theory and graph neural network technology, the advantages of them in processing heterogeneous data of recommendation system are clarified. The designed model integrates node embeddedness learning, neighbor aggregation and attention mechanism to capture node attributes effectively, mine semantic associations and potential patterns in heterogeneous graphs, and maintain node type differences. By learning node embeddedness such as user and item, the model reveals the user-item interaction rule and cross-node type influence, and uses attention mechanism to optimize recommendation accuracy and explainability. Publicly accessible heterogeneous datasets serve as the experimental platform to demonstrate that the proposed HGNN-based recommendation model surpasses prevailing algorithms across multiple performance metrics, namely accuracy, diversity, and coverage. Ablation studies substantiate the efficacy of the model's constituent parts and the judicious exploitation of heterogeneous information. Concluding the investigation, the model's scalability and real-world applicability are deliberated upon.