Improving healthy food recommender systems through heterogeneous hypergraph learning

Recommender systems in health-conscious recipe suggestions have evolved rapidly, particularly with the integration of both homogeneous and heterogeneous graphs. However, incorporating IoT devices into healthcare, such as wearable fitness trackers and smart nutrition scales, presents new challenges. These devices generate vast amounts of dynamic, personalized data, which traditional Graph Neural Network (GNN) models — limited to simple pairwise connections — fail to capture effectively. For example, IoT sensors tracking daily nutrient intake require complex, multi-faceted analysis that traditional methods struggle to handle. To overcome these limitations, researchers have employed hypergraphs, which capture higher-order relationships among nodes, such as user–food and ingredient interactions. Traditional methods using static weights in the Laplacian hypergraph, inspired by homogeneous graph techniques, often fail to account for users’ evolving health interests. Our study introduces a novel approach for recommending healthy foods by leveraging user–food and food-ingredient hyperedges, integrating both convolution and attention-based hypergraph mechanisms to dynamically adjust weights based on user similarities. Unlike previous methods, we convert the heterogeneous hypergraph into a homogeneous space, using a unified loss function to generate recommendations that adapt to individual users’ changing dietary preferences. The model is evaluated on five metrics — AUC, NDCG, Precision, Recall, and F1-score — and shows superior performance compared to existing models on two real-world food datasets, Allrecipes and Food.com. Our results demonstrate significant improvements in recommendation accuracy and personalization, showcasing the system’s effectiveness in integrating IoT data for more responsive, health-focused food suggestions.