Adaptive dynamic hypergraph learning for ingredient aware food recommendation

Food recommendation systems face fundamental challenges in modeling the complex, compositional relationships among users, foods, and ingredients. Traditional collaborative filtering and Graph Neural Networks rely on pairwise connections that oversimplify culinary interactions, while existing hypergraph approaches use static weights that fail to adapt to dynamic user preferences and ingredient semantics. We propose FRMADHG (Food Recommendation with Multi-objective Adaptive Dynamic Hypergraph), a novel framework that captures higher-order interactions through a tripartite hypergraph where foods serve as hyperedges connecting users and ingredients. Our key innovations include: (1) a complexity-aware adaptive attention mechanism that dynamically switches between efficient cosine similarity and sophisticated learnable attention based on local interaction complexity, (2) type-specific embedding propagation rules that respect the distinct semantic roles of users, foods, and ingredients, (3) dynamic Laplacian construction that evolves during training to emphasize semantically coherent relationships, and (4) a multi-objective learning strategy combining triplet ranking, contrastive learning, and regularization for robust embeddings. FRMADHG provides multi-granular explainability through ingredient-level importance scores and path-based reasoning, enabling transparent dietary decisions. Comprehensive experiments on two large-scale datasets-Food.com (32,000 users, 18,500 foods, 1,270 unique ingredients) and Allrecipes (25,400 users, 16,200 foods, 1,050 ingredients)-demonstrate significant improvements over state-of-the-art methods: 19.8% relative gain in Precision@10 compared to collaborative filtering baselines, 12.4% improvement over recent hypergraph approaches, and 18.7% enhancement in Recall@10 (all with [Formula: see text]). Ablation studies validate the critical contributions of dynamic hypergraph construction (12.4% performance gain), adaptive attention mechanism (11.8%), and contrastive learning (9.7%), while user studies confirm the effectiveness of our ingredient-level explanations for building user trust and satisfaction.