This is the code for "Community-Invariant Graph Contrastive Learning" (CI-GCL). CI-GCL adopt learnable data augmentation with Community-Invariant constraint on both topology and features. And all these parts are jointly optimized to make sure the augmentation schemes can benefit from contrastive loss, CI constraints and downstream classifiers.

Code is tested in Python 3.10.13. The requirements are listed below:

pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118
pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu118.html
pip install dgl
pip install networkx
pip install tqdm tabulate
pip install scipy
pip install scikit-learn

For different graph-level tasks, we provides:

python graph_classification.py --dataset_name ${dataset_name} --gpu ${gpu}

${dataset_name} is the dataset name (please refer to TU dataset), ${gpu} is the lanched GPU ID

python graph_transfer_pretrain_and_finetune.py --dataset_name ${dataset_name} --ft_dataset ${ft_dataset}

where ${dataset_name} is the pretrain dataset, and ${ft_dataset} is the finetune dataset

We randomly drop and add [0.05, 0.1, 0.15, 0.2, 0.25, 0.3] edges in topology robustness test. And randomly add gaussian noise to [0.05, 0.1, 0.15, 0.2, 0.25, 0.3] positions in feature robustness test.

python robustness_check.py --dataset_name ${dataset_name}
python robustness_check_feature.py --dataset_name ${dataset_name}