Comprehensive Analysis of GNN Approaches for Food Recipe Recommendation

Environment Setup

1. Setup python virtual environement with required depedencies:

   sudo chmod +x install_requirements.sh
   
   bash install_requirements.sh
   
   source venv/bin/activate
   

2. Set environment variables

   cp .env.example .env
   

Graph Dataset Generation

In order to generate graph datasets, you can run the following command:

python3 -m graph_generation.graph_dataset_generator

The details can be found in the graph_dataset_generator.py

The graph generation contains the following steps:

1. Downloading data:
   We download raw Food.com Recipes and Interactions files (shuyangli94/food-com-recipes-and-user-interactions) via KaggleAPI.
   Please note KAGGLE_USERNAME and KAGGLE_KEY needs to be set in the .env file.

2. Processing users:
   In order to generate user nodes, we only retrieve their IDs from RAW_interactions.csv as they do not contain rich properties.
   Notes:

   • Originally, we intented to consider techniques as a property for user ndoes, but they seem to be based on all the edges (interactions). Considering we split the edges into training, validation, and test sets, adding techniques might lead to biased evaluation results.

   • RAW_interactions.csv contains all the users, hence there are no unseen users for validation and testing.

3. Processing recipes:
   In order to generate recipes nodes, we use RAW_recipes.csv to process all the following selected properties with proper encoders:

   • name: SequenceEncoder
   • steps: SequenceListEncoder
   • tags: SequenceListEncoder
   • ingredients: SequenceListEncoder
   • nutrition: ListEncoder
   • n_steps: IdentityEncoder
   • minutes: IdentityEncoder

   In generall, the encoders make sure the values are in numeric form with correect data types and also convert them into Torch tensors to prepare them for training and inference.

4. Processing edges:
   In this step, we process edges with a single property (rating) using RAW_interactions.csv dataset.

5. Generating graph:
   In this step, we generate a heterogeneous graph dataset given the processed users, recipes and their respective edges. We construct heterogeneous graphs using torch_geometric.data.HeteroData since we have two different types of nodes (users and edges).

6. Generating data splits: After generating the graph, we split them into three graph with distinct edges using torch_geometric.transforms.RandomLinkSplit:

   • Train graph: It contains 85% of all edges, 20% percent of which is used as supervision edges.
   • Validation graph: It contains 5% of the remaining edges.
   • Test graph: It contains 10% of the remaining edges.

   These are the hyperparameters in the code that can control the edge ratios for each split:

   supervision_train_ratio = 0.2
   validation_ratio = 0.05
   test_ratio = 0.1
   

7. Storing graphs:
   Finally, we store the generated graphs in PyTorch (.pt) files which can be used for training, validation and testing later.