Certified Graph Unlearning

This is the repository of certified (approximate) machine unlearning for simplified graph convolutional networks (SGCs) with theoretical guarantees. Our accompany papers are as follows:

• [ICLR 2023] Efficient Model Updates for Approximate Unlearning of Graph-Structured Data
• [NeurIPS 2022 GLFrontier workshop] Certified Graph Unlearning

For our extension to graph classification with graph scattering transform, please check the repository and the accompany paper.

• [TheWebConf 2023] Unlearning Nonlinear Graph Classifiers in the Limited Training Data Regime

Package Information

pytorch=1.10.0
torch-geometric=2.0.4
ogb

Example

Our method supports three types of removing requests that could possibly arise in graph unlearning: node feature unlearning, node unlearning, and edge unlearning. We also include the comparison with complete retraining and one method designed for unstructured unlearning by Guo et al.

Here are three examples for Cora dataset. Detailed description of each input can be found in the code.

Node feature unlearning

python sgc_feature_node_unlearn.py --dataset='cora' --std=0.1 --num_removes=800 --train_mode='ovr' --prop_step=2 --lr=0.5 --removal_mode='feature' \
                                   --disp=100 --trails=1 --compare_retrain --compare_guo --optimizer='LBFGS' --data_dir='../PyG_datasets'

Node unlearning

python sgc_feature_node_unlearn.py --dataset='cora' --std=0.1 --num_removes=800 --train_mode='ovr' --prop_step=2 --lr=0.5 --removal_mode='node' \
                                   --disp=100 --trails=1 --compare_retrain --compare_guo --optimizer='LBFGS' --data_dir='../PyG_datasets'

Edge unlearning

python sgc_edge_unlearn.py --dataset='cora' --std=0.1 --num_removes=2000 --train_mode='ovr' --prop_step=2 --lr=0.5 --removal_mode='edge' \
                           --disp=100 --trails=1 --compare_retrain --optimizer='LBFGS' --data_dir='../PyG_datasets'

Contact

Please contact Chao Pan (chaopan2@illinois.edu), Eli Chien (ichien3@illinois.edu) if you have any question.

Citation

If you find our code or work useful, please consider citing our paper:

@inproceedings{
chien2023efficient,
title={Efficient Model Updates for Approximate Unlearning of Graph-Structured Data},
author={Eli Chien and Chao Pan and Olgica Milenkovic},
booktitle={International Conference on Learning Representations},
year={2023},
url={https://openreview.net/forum?id=fhcu4FBLciL}
}

@inproceedings{
chien2022certified,
title={Certified Graph Unlearning},
author={Eli Chien and Chao Pan and Olgica Milenkovic},
booktitle={NeurIPS 2022 Workshop: New Frontiers in Graph Learning},
year={2022},
url={https://openreview.net/forum?id=wCxlGc9ZCwi}
}