UIFRS-HAN: User interests-aware food recommender system based on the heterogeneous attention network

In recent years, the surge in social media platform usage has sparked a heightened interest in applying recommender systems (RSs) within the food industry. Traditionally, the exploration of user interests focused on analyzing behaviors linked to food selection. The availability of user interaction datasets now provides avenues for deeper insights into food content and intricate user relationships. This paper advocates strategically integrating Heterogeneous Information Networks (HIN) into recommender system frameworks. It introduces the Heterogeneous Attention Network-based User Interests-Aware Food Recommender System (UIFRS-HAN), designed for personalized food recommendations. By leveraging HIN and a two-step attention mechanism, UIFRS-HAN captures diverse entities and relationships within a unified framework. UIFRS-HAN employs an attention technique to reconstruct node features and edges, incorporating a dual hierarchical attention mechanism for improved unsupervised learning of attributed graph representations. Besides, HIN allows the model to uncover meaningful relationships between nodes, particularly when directed relationships are unclear. Through a defined meta-path-based attention mechanism, UIFRS-HAN generates diverse recommendations based on users’ interests across various relations among different types of nodes of the HIN. By discerning intricate patterns and correlations, UIFRS-HAN surpasses traditional approaches in delivering refined and contextually relevant recommendations. The proposed model enhances representation depth and accuracy by employing node embedding through a hierarchical meta-path structure. Rigorous testing on Allrecipes.com and Food.com datasets, compared against 15 baselines and state-of-the-art models, confirms the technical soundness and superiority of UIFRS-HAN in providing precise and personalized food recommendations. • A novel food recommender system based on a heterogeneous attention network. • The dual attention method is used to learn the meta-path in HIN. • Employing unsupervised learning based on hierarchical attention in the HIN. • The experiment method involves two real datasets based on heterogeneous graphs.