CodeAppendix

Dependencies

• pytorch 1.5.0
• pytorch-lightning 0.7.3
• torch-geometric 1.4.3
• numpy 1.18.1
• scipy 1.5.1
• wandb 0.9.4 (please run wandb off to disable the sync to wandb server)

File Structure

• data: Directory for dataset.
• model:
  • node.py: The node module of EGAT.
  • edge.py: The edge module of EGAT.
  • mgcn.py: The edge and node modules of MGCN, including the EGAT_MGCN (AttentionVertexModule)
  • nnconv.py: The node module of NNConv, including the EGAT_NNConv (AttentionNNConv)
  • net.py: The network structure of EGAT, for both AMLSim (AMLSimNet) and citation networks (Cora, Citeseer and PubMed) (CitationNet). The structure of CitationNet is hard coded.
• trainer: The training process (see: pytorch-lightning) of AMLSim and citation networks.
• transforms: The transformers of dataset.
• dataset.py: Some of the preprocessing of AMLSim and all the preprocessing of citation networks.
• main.py: The entry file.
• config.yml: Hyperparameter config file.

Usage

Dataset Prepare

Please copy all dataset to data directory. (available at this url)

Hyperparameters

You can control the hyperparameter in config.yml. where the meaning of each hyperparameter is commented .

Train

Run python main.py to train the model. The results are reported in the terminal.