Why Does Dropping Edges Usually Outperform Adding Edges in Graph Contrastive Learning? (AAAI 2025)

This is our PyTorch implementation of EPAGCL:

Yanchen Xu⁺, Siqi Huang⁺, Hongyuan Zhang^*, and Xuelong Li^*, "Why Does Dropping Edges Usually Outperform Adding Edges in Graph Contrastive Learning?", in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. (arXiv)

The main script is main.py used for training on various datasets and evaluating periodically during training. Evaluation is triggered every args.eval epochs and will not back-propagate any gradient to the encoder.

Requirements

This code package was developed and tested with Python 3.11.7. The required dependencies are as follows:

• numpy 1.26.4
• ogb 1.3.6
• scipy 1.11.4
• thop 0.1.1.post2209072238
• torch 2.3.0
• torch_geometric 2.5.3
• tqdm 4.65.0

How to run

To run EPAGCL:

python main.py --add_single

To run EPAGCL that computes loss in batch to reduce memory burden (batch size = 256 in defalut and can be changed in line 21 of Model.py):

python main.py --add_single --batch_compute

To compute EPR of certain dataset (set in line 8 of EPR.py):

python EPR.py

To conduct Wilcoxon signed-rank test (set accuracy in line 2 and line 3 of wilcoxon.py):

python wilcoxon.py

Hyper-parameters

Hyper-parameters used for each dataset is listed in the appendix of our paper. To run EPAGCL with different setting, one may change the corresponding variables in Arguments.py or set them through command line.

For example, to run EPAGCL on ogbn-arxiv:

python main.py --add_single --dataset ogbn-arxiv --edge_drop_rate_1 0.6 --edge_drop_rate_2 0.6 --batch_compute

All hyper-parameters set through Arguments.py will be printed in args.txt.

Ablation Settings

There are some options as follows for ablation study:

• --add_single: add edges for only one view.
• --not_add_edge: not add edges for generating views.
• --not_drop_edge: not drop edges for generating views.
• --add_edge_random: add edges randomly for generating views.

These settings will be printed in result.txt for ease of viewing.

Contact

If you have any questions, please email Yanchen Xu (primary author of code, pkuxuyanchen@163.com) or Hongyuan Zhang (hyzhang98@gmail.com).

Citation

If you find the code useful for your research, please consider citing our work:

@inproceedings{EPAGCL,
  author={Xu, Yanchen and Huang, Siqi and Zhang, Hongyuan and Li, Xuelong},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
  title={Why Does Dropping Edges Usually Outperform Adding Edges in Graph Contrastive Learning?}, 
  year={2025},
  pages={},
}