This repository contains the official code base of the paper Cooperative Graph Neural Networks, accepted to ICML 2024

To reproduce the results please use Python 3.9, PyTorch version 2.0.0, Cuda 11.8, PyG version 2.3.0, and torchmetrics.

pip install torch==2.0.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.0.0+cu118.html
pip install torch-geometric==2.3.0
pip install torchmetrics ogb rdkit
pip install matplotlib

The synthetic datasets RootNeighbours and Cycles should be generated with a seed of 0.

Available in Pytorch Geometric.

The TUDatasets and Social network graph classification datasets can be found at datasets/.

Make sure to unzip the REDDIT-MULTI-5K.zip first!

unzip datasets/REDDIT-MULTI-5K.zip -d datasets

The Peptides-func dataset can be found at datasets/peptides-functional.

The script we use to run the experiments is ./main.py. Note that the script should be run with . as the main directory or source root.

The parameters of the script are:

• --dataset: name of the dataset.
• --pool: name of the graph pooling.
• --learn_temp: a flag to be used when the temperature is learned.
• --temp_model_type: the type of GNN that learns a temperature for the Gumbel-softmax estimator (relevant only if learn_temp is present).
• --tau0: the tau0 parameter in a learnable temperature model (relevant only if learn_temp is present).
• --temp: the temperature of the Gumbel-softmax estimator (relevant only if learn_temp is not present).
• --max_epochs: the number of epochs.
• --batch_size: the batch size.
• --lr: the learn rate.
• --env_model_type: the type of GNN the environment network uses.
• --env_num_layers: the environment network's number of layers.
• --env_dim: the environment network's hidden dimension.
• --skip: a flag that is used to include skip connections.
• --batch_norm: the batch size.
• --layer_norm: a flag that is used to include layer_norm.
• --dec_num_layers: the number of layers the decoder MLP uses.
• --pos_enc: the type of positional encoding used.
• --act_model_type: the type of GNN the action network uses.
• --act_num_layers: the action network's number of layers.
• --act_dim: the action network's hidden dimension.
• --seed: a seed to set random processes.
• --gpu: the number of the gpu that is used to run the code on.
• --fold: a specific fold of the dataset (only applicable to a portion of the datasets used).
• --weight_decay: the weight decay.
• --step_size: the step_size of the StepLR scheduler (only applicable to a portion of the datasets used).
• --gamma: the gamma of the StepLR scheduler (only applicable to a portion of the datasets used).
• --num_warmup_epochs: the num_warmup_epochs of cosine_with_warmup_scheduler (only applicable to a portion of the datasets used).

To perform experiments with a CoGNN($\mu, \mu$) model with 3 environment layers and an environment hidden dimension of 64, with a 1-layer action network with a hidden dimension of 16. See an example for the use of the following command:

python -u main.py --dataset roman_empire --env_model_type MEAN_GNN --act_model_type MEAN_GNN --env_dim 64 --env_num_layers 3 --act_dim 16 --act_num_layers 1 --seed 0

If you make use of this code, or its accompanying paper, please cite this work as follows:

@inproceedings{finkelshtein2023cooperative,
  title = "Cooperative Graph Neural Networks",
  author = "Ben Finkelshtein and Xingyue Huang and Michael Bronstein and {\.I}smail {\.I}lkan Ceylan",
  year = "2024",
  booktitle = "Proceedings of Forty-first International Conference on Machine Learning (ICML)",
  url = "https://arxiv.org/abs/2310.01267",
}