Code Repository for Self-Enhanced GNN

This is the code repository for Self-Enhanced GNN: Improving Graph Neural Networks Using Model Outputs (IJCNN 2021).

Dependencies

CUDA==10.2.89
python==3.6.9
torch==1.5.0
torch_geometric==1.4.3

prepare

Run split.py to create random train/val/test splits for each dataset.

topology update

Run topology_update.py to create updated edges and save them.

Then run topology_update_train.py to train the models and save the results.

Use stats.py --mode val_test_result to analyze the results.

training node augmentation

Run training_node_augmentation.py to train with TNA algorithm. Remember to save the outputs to a text file.

Use tna_parser.py to analyze the above generated outputs.

Run tna_ensemble.py and tna_distilled.py to get the ensemble and distillation results, respectively.

combined

Run combined_models.py to run the combined version. The topology update result filename and the training node augmentation outputs template needs to be specified to get the corresponding threshold when running this script.

Bibtex

@inproceedings{yang2021selfenhanced,
    title={Self-Enhanced GNN: Improving Graph Neural Networks Using Model Outputs}, 
    author={Han Yang and Xiao Yan and Xinyan Dai and Yongqiang Chen and James Cheng},
    year={2021},
    booktitle={International Joint Conference on Neural Networks}
}