CSGCL: Community-Strength-Enhanced Graph Contrastive Learning

PyTorch implementation for IJCAI 2023 Main Track Paper "CSGCL: Community-Strength-Enhanced Graph Contrastive Learning" (https://www.ijcai.org/proceedings/2023/0229.pdf).

The code is based on the implementation of GCA.

We also have a Chinese introduction blog on Zhihu.

Dependencies

• Python 3.8.8
• PyTorch 1.8.1
• torch_geometric 2.0.1
• cdlib 0.2.6
• networkx 2.5.1
• numpy 1.22.4

Quick Start

The best hyperparameters for node classification (as reported in Appendix C.2 of the paper) can be found in ./param, which will be directly loaded by --param:

python train.py --dataset WikiCS --param local:wikics.json

You can change the parameter by either .json files (NOT RECOMMENDED) or simply add it to the command, for example:

python train.py --dataset WikiCS --param local:wikics.json --num_epochs 5000

Results

All experiments are conducted on an 11GB NVIDIA GeForce GTX 1080 Ti GPU with CUDA 11.3. The node classification results are shown below.

	Wiki-CS	Computers	Photo	Coauthor-CS
GCA (best conf)	78.20±0.04	87.99±0.13	92.06±0.27	92.81±0.19
CSGCL	78.60±0.13	90.17±0.17	93.32±0.21	93.55±0.12

Citing

Please cite our paper for your research if it helps:

@article{csgcl,
  title={CSGCL: Community-Strength-Enhanced Graph Contrastive Learning}, 
  author={Han, Chen and Ziwen, Zhao and Yuhua, Li and Yixiong, Zou and Ruixuan, Li and Rui, Zhang},
  journal={CoRR},
  volume={abs/2305.04658},
  year={2023}
}