joyce04/graph_agu

Data Augmentation Approaches for Graph Neural Networks

Highlights

• Review recent data augmentation approaches for GNNs
• Experiments with following baseline models: GCN, GAT, GraphSage
• Data augmentation studied:
  • Node-level augmentation : FLAG[1] git repository
  • Topology-level augmentation : G-Aug[2] git repository
  • Topology-level updates : DropEdge[3] git repository

optuna is used to find best parameters

Requirements

• Python 3.8.2
• optuna==2.7.0
• torch==1.8.0
• torch-geometric==1.6.3
• torch-scatter==2.0.6
• torch-sparse==0.6.9 Please refer to the requirements.txt for any other packages.

Implementation Details

All GNNs are implemented with pytorch-geometric.

Usage

To run the demo :

• For baseline gcn, graphsage, gat : sh ./script/base_trial.sh
• For FLAG gcn, graphsage, gat : sh ./script/flag_trial.sh
• For DropEdge gcn, graphsage, gat : sh ./script/de_trial.sh
• For G-Aug gcn, graphsage, gat : sh ./script/de_trial.sh Please modify json files in ./configs to change any hyperparameters.

References

[1] Kong, Kezhi, et al. "Flag: Adversarial data augmentation for graph neural networks." arXiv preprint arXiv:2010.09891 (2020).
[2] Zhao, Tong, et al. "Data Augmentation for Graph Neural Networks." arXiv preprint arXiv:2006.06830 (2020). [3] Rong, Yu, et al. "Dropedge: Towards deep graph convolutional networks on node classification." arXiv preprint arXiv:1907.10903 (2019).