Self-Supervised Graph Representation Learning via Topology Transformations

This repository is the official PyTorch implementation of the following paper:

Xiang Gao, Wei Hu, Guo-Jun Qi, "Self-Supervised Graph Representation Learning via Topology Transformations," IEEE Transactions on Knowledge and Data Engineering (TKDE), December 2021.

Requirements

• Python3>=3.7.10
• pytorch>=1.9.0
• tensorboardX>=1.9
• torch_geometric>=1.7.2

Note: We are not sure whether the code can be run properly by using a lower version of the relevant package.

Usage

Tip: Experimental results of graphs usually face greater randomness than images. We suggest you run the experiment more than one time and select the best result.

Unsupervised Training

To train a feature extractor in an unsupervised fashion, run

python main.py --phase backbone --dataset cora --perturbation-rate 0.7 --hidden-channels 512 --k 2 --num-epochs 512 --lr 0.0001 --use-cuda true --device 0 --save-dir ./results

Supervised Evaluation

After training the feature extractor, you need to train the classifier by running the following command:

python main.py --phase classifier --dataset cora --perturbation-rate 0.7 --hidden-channels 512 --k 2 --backbone ./results/cora_best.pt --lr 0.001 --use-cuda true --device 0 --save-dir ./results

Reference

Please cite our paper if you use any part of the code from this repository:

@article{gao2021topoter,
  title={Self-Supervised Graph Representation Learning via Topology Transformations},
  author={Gao, Xiang and Hu, Wei and Qi, Guo-Jun},
  journal={IEEE Transactions on Knowledge and Data Engineering (TKDE)},
  month={December},
  year={2021}
}

Acknowledgement

Our code is released under MIT License (see LICENSE for details). Some of the code in this repository was borrowed from the following repositories:

• DGI
• GMI