Home

Awesome

GraphAT

Code for the GraphAT, GraphVAT, and GCN-VAT in our paper "Graph Adversarial Training: Dynamically Regularizing Based on Graph Structure", [paper].

Environment

Python 3.6.1 :: Continuum Analytics, Inc.

tensorflow 1.8.0

numpy 1.18.1

Reproduction

Once configured the required environment, the prediction performance reported in our paper can be reproduced by running the following commands (Table 4).

GraphAT

python gvat_citation.py --gat_loss=True --num_neighbors 2 --epsilon_graph 0.01 --beta 1.0 --dropout 0.0 --dataset cora --early_stopping 10
python gvat_citation.py --gat_loss=True --num_neighbors 2 --epsilon_graph 0.01 --beta 0.5 --dropout 0.0 --dataset citeseer --early_stopping 10

GraphVAT

python gvat_citation.py --gat_loss=True --vat_loss=True --epsilon 1.0 --alpha 0.5 --xi 1e-05 --num_neighbors 2 --epsilon_graph 0.01 --beta 1.0 --dropout 0.0 --dataset cora --early_stopping 10
python gvat_citation.py --gat_loss=True --vat_loss=True --epsilon 1.0 --alpha 0.5 --xi 1e-06 --num_neighbors 2 --epsilon_graph 0.01 --beta 0.5 --dropout 0.0 --dataset citeseer --early_stopping 1

GCN-VAT

python vat_citation.py --epsilon 0.01 --alpha 1.0 --xi 0.001 --dropout 0.0 --dataset cora --early_stopping 10
python vat_citation.py --epsilon 0.05 --alpha 0.5 --xi 0.0001 --dropout 0.0 --dataset citeseer --early_stopping 10

Cite

If you use the code, please kindly cite the following paper:

@article{feng2019graph,
  title={Graph adversarial training: Dynamically regularizing based on graph structure},
  author={Feng, Fuli and He, Xiangnan and Tang, Jie and Chua, Tat-Seng},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2019},
  publisher={IEEE}
}

Contact

fulifeng93@gmail.com