Paper | Poster | Supplementary

The PyTorch implementation of Directed Graph Contrastive Learning (DiGCL).

In this paper, we present the first contrastive learning framework for learning directed graph representation.

If you find our work useful, please considering citing

@article{tong2021directed,
  title={Directed Graph Contrastive Learning},
  author={Tong, Zekun and Liang, Yuxuan and Ding, Henghui and Dai, Yongxing and Li, Xinke and Wang, Changhu},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}

Our project is developed using Python 3.7, PyTorch 1.7.0 with CUDA10.2. We recommend you to use anaconda for dependency configuration.

First create an anaconda environment called DiGCL by

conda create -n DiGCL python=3.7
conda activate DiGCL

Then, you need to install torch manually to fit in with your server environment (e.g. CUDA version). For the torch and torchvision used in my project, run

conda install pytorch==1.7.0 torchvision==0.6.0 cudatoolkit=10.2 -c pytorch

Besides, torch-scatter and torch-sparse are required for dealing with sparse graph. For these two packages, please follow their official instruction torch-scatter and torch-sparse.

Other requirements can be set up through:

cd DiGCL
pip install -e .

cd code
python train_digcl.py --gpu_id 0 --dataset cora_ml --curr-type log
python train_digcl.py --gpu_id 0 --dataset citeseer

The --dataset argument can be one of [cora_ml, citeseer] and the --curr-type argument can be one of [linear, log, exp, fixed].

DiGCL is released under the MIT License. See the LICENSE file for more details.

We are grateful for the following enlightening works, which are also of great use in our work.

• Graph Contrastive Learning Library for PyTorch: PyGCL
• Graph Contrastive Learning with Adaptive Augmentation: GRACE and GCA
• Graph Contrastive Learning with Augmentations: GraphCL
• Our another supervised approach to process directed graphs: DiGCN
• MagNet: A Neural Network for Directed Graphs: MagNet

The template is borrowed from Pytorch-Geometric benchmark suite. We thank the authors of following works for opening source their excellent codes. Pytorch-Geometric, Graph2Gauss, and GNN-benchmark.