Social recommendation system based on heterogeneous graph attention networks

Heterogeneous graph attention networks are becoming a popular choice, particularly in areas like social recommendation systems, surpassing conventional graph attention networks. Our research highlights their adaptability to the complexity of relationships among different types of nodes, effectively addressing the 'cold start' challenge by leveraging a variety of information. This paper introduces an innovative social recommendation system based on deep heterogeneous graph attention networks to produce embeddings that represent users and items from different perspectives. By incorporating contextual information and user behavior patterns, our approach enhances prediction accuracy. Additionally, we employ a multiclass classification method to predict user ratings for unrated items, providing a comprehensive recommendation solution. Experimental results, evaluated on three benchmark datasets, show substantial improvements, with HR@k increasing by 23.55-44.68% and NDCG@k rising by 21.64-27%. These enhancements reflect significant gains in accuracy, relevance, and ranking performance compared to baseline models. Thus, heterogeneous graph attention networks emerge as a preferred approach to enhance the performance of social recommendation systems, offering a holistic solution to address diverse challenges in recommendation scenarios.