Health-Aware Contrastive Learning for Recipe Recommendation

This paper addresses the shortcomings of traditional recipe recommendation systems, which often neglect nutritional balance and health needs, while facing challenges such as long-tail effects and data noise. We propose the "Health-Aware Contrastive Learning for Recipe Recommendation" (HACL) model. HACL constructs a nutritional assessment framework to align with users' health requirements and incorporates contrastive learning strategies to address the long-tail distribution and noise in knowledge graphs (KG). By building local and non-local knowledge graphs and implementing multi-level contrastive learning, HACL effectively extracts intrinsic graph information and optimizes long-tail and noise issues. Additionally, the model integrates contrastive loss for health attributes during the learning process, emphasizing the importance of health factors and enhancing the health orientation of recommendations. Experimental validation on two real recipe datasets demonstrates the effectiveness and superiority of HACL.