Heterogeneous Graph Neural Networks in Representation Learning for Sustainable Proteins: An Innovative Solution to the Cold Start Problem

This article introduces a heterogeneous graph neural network architecture for mapping novel sustainable proteins (such as mycelium protein, precision-fermented casein, and microalgae protein) into the culinary space, addressing the cold start problem caused by their lack of historical recipe data. The core idea is to use multi-modal features such as flavor, nutritional components, and processing characteristics to model ingredient relationships via heterogeneous graphs, and combine supervised learning with contrastive learning to enhance representation capabilities, providing references for the culinary applications of novel proteins.

With the global growth in demand for sustainable food, novel proteins such as mycelium protein, precision-fermented casein, and microalgae protein have gained attention. However, these ingredients lack historical recipe data, making it difficult for traditional recommendation systems to integrate them into existing culinary systems, leading to the "cold start" problem: new ingredients cannot be effectively recommended due to the absence of interaction data, hindering acceptance by chefs and consumers.

The project constructs a heterogeneous graph containing five types of nodes: ingredients, flavors, nutrition, processing, and cuisines. Edge relationships include has_flavour, contains, prepared_by, belongs_to, etc. The model uses a single-layer HeteroAttentionNet: mean aggregation of messages for the same edge type, attention pooling across edge types; dual loss strategy (cross-entropy supervised loss for predicting cuisine labels + InfoNCE contrastive loss to enhance robustness).

During training, novel proteins are fully retained (not involved in loss calculation, only feature edge connections are kept). In the inference phase, messages are propagated through feature edges to calculate similarity with known ingredients and return neighbors and cuisines. Examples: Precision-fermented casein's neighbors are halloumi cheese and chickpeas (Middle Eastern cuisine); Mycelium Protein X's neighbors are tofu and miso (Japanese cuisine).

On the synthetic dataset (25 labeled ingredients, 9 cuisines), after 200 epochs, the supervised loss decreased from 2.219 to 0.045, and the contrastive loss decreased from 7.918 to 1.147; the leave-one-out Top3 cuisine recall rate was 0.920 (confidence interval 0.800-1.000). The implementation is a pure PyTorch CPU version with a clear code structure, and a fixed seed=42 ensures reproducibility.

Limitations: Synthetic data not validated, small dataset size, loss balance to be optimized, single-layer message passing limits interactions. Future directions: Integrate real data sources (FlavorDB, USDA), adjust loss weights, increase network depth, introduce more modalities such as texture/appearance.

The project accelerates the adoption of sustainable foods: provides culinary inspiration for chefs, guides product positioning, and supports personalized recommendations. Core contributions: multi-modal heterogeneous graph design, cold start evaluation mechanism, combination of supervised and contrastive learning. Although based on synthetic data, the architecture is clear and easy to extend, providing a reference for cross-application of food technology and GNN.