rdangovs / 6883-project

FLAG: Adversarial Data Augmentation for Graph Neural Networks

This is the official repo for the paper FLAG: Adversarial Data Augmentation for Graph Neural Networks.

TL;DR: FLAG augments node features to generalize GNNs on both node and graph classification tasks.

Highlights

• Simple, adding just a dozen lines of code
• General, directly applied to any GNN baseline
• Versatile, working on both node and graph classification tasks
• Scalable, minimum extra memory overhead, working on the original infrastructure

Experiments

To reproduce experimental results for DeeperGCN, visit here.

Other baselines including GCN, GraphSAGE, GAT, GIN, MLP, etc. are available here.

To view the empirical performance of FLAG, please visit the Open Graph Benchmark Node an Graph classification leaderboards.

Requirements

• ogb=1.2.3
• torch-geometric=1.6.1
• torch=1.5.0

Citing FLAG

If you use FLAG in your work, please cite our paper.

@misc{kong2020flag,
      title={FLAG: Adversarial Data Augmentation for Graph Neural Networks}, 
      author={Kezhi Kong and Guohao Li and Mucong Ding and Zuxuan Wu and Chen Zhu and Bernard Ghanem and Gavin Taylor and Tom Goldstein},
      year={2020},
      eprint={2010.09891},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}