LPM93/italian-cuisine-gat

Italian Regional Cuisine Classification - GNN Models

This repository contains the trained Graph Attention Network (GAT) models for classifying Italian recipes by region.

Models

This repository contains three approaches:

1. Fine-grained (20 regions): models/fine_grained/best_model.pt

   • Accuracy: 20.3%
   • Classifies recipes into 20 Italian regions

2. Macro-region (4 classes): models/macro_region/best_model_macro.pt

   • Accuracy: 59.5%
   • Classifies into North, Center, South, Islands

3. Hierarchical: models/hierarchical/{region}/best_model.pt

   • Two-level classifier (macro → fine)
   • Accuracy: 22.3%

Usage

Download models using the companion code repository:

# Clone the main repository
git clone https://github.com/LeonardoPaccianiMori/cuisine-project-temp.git
cd cuisine-project-temp

# Download models from Hugging Face
python scripts/download_models.py

Model Architecture

• Type: Heterogeneous Graph Attention Network (HeteroGAT)
• Framework: PyTorch Geometric
• Graph structure: Recipe → Ingredient, Recipe → Step edges
• Hidden dimension: 256
• Attention heads: 4
• Dropout: 0.3

Citation

If you use these models, please cite:

@misc{italian-cuisine-gnn,
  author = {Leonardo Pacciani-Mori},
  title = {Italian Regional Cuisine Classification with Graph Neural Networks},
  year = {2025},
  url = {https://github.com/LeonardoPaccianiMori/portfolio-italian-cuisine}
}

License

MIT License - See the main repository for details.

Data sources:

• Artusi recipes (public domain, 1891)
• AIC recipes (consult https://www.cucinaitaliana.it for usage terms)