/WSDM23-GSR

Primary LanguageJupyter Notebook

WSDM23-GSR

An implementation of the GSR model proposed in Self-Supervised Graph Structure Refinement for Graph Neural Network in WSDM23. Please kindly note that this work is inspired by the great work GAuG, especially the adding and removing edges part.

Requirements

  • Python >=3.8.5
  • Pytorch == 1.8.1
  • DGL == 0.6.2
  • SciPy == 1.6.2
  • NetworkX==2.5.1

Usage

  • Step 1 : Unzip the data.zip file
  • Step 2: Run the experiments using the command below

to recover our results, specify "-l" option

python root_to_src_dir/src/models/GSR/trainGSR.py -dcora -l

Usage

Experiment Settings

  • dataset: The dataset to run on.
  • gpu: GPU id to use.
  • train_percentage:The train ratio, 0 for default split.
  • load_default_config: Whether load default config or use parsed config.
  • early_stop: Number of epoch for early stop.
  • seed: Training seed.

Model Settings

  • epochs: The maximum epoch to train in the fine-tuning process.
  • intra_weight: The weight of intra contrastive loss, $\alpha$ in the paper.
  • fsim_weight: The weight of feature similarity in estimating edge probability, $\beta$ in the paper.
  • add_ratio: Percentage of non-existing edge to add in the graph refinement process.
  • rm_ratio: Percentage of existing edge to remove in the graph refinement process.
  • fan_out: Number of neighbors to sample in first and second order subgraphs.
  • p_epochs: Number of pre-training epochs.
  • p_batch_size: The pre-training batch size.
  • prt_lr: The pre-training learning rate.

Citation

If you find our work useful, please consider citing our work:

@inproceeding{zhao2023gsr,
  title={Self-Supervised Graph Structure Refinement for Graph Neural Networks},
  author = {Jianan Zhao and Qianlong Wen and Mingxuan Ju and Chuxu Zhang and Yanfang Ye},
  booktitle = {WSDM},
  year      = {2023},
}