A powerful and efficient method for graph unlearning when the size of training dataset is limited. Our accompany paper is as follows:

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

This work is developed based on our prior work certified graph unlearning. Feel free to check this repository and the following accompany papers for more details.

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

pytorch
scikit-learn
pytorch-dp
pytorch-geometric
tqdm

We assume the following project directory structure in our code:

<root>/
--> data/
--> results/
--> scripts/

If you have a different path for the datasets, you might need to change the utility function in datasets.py.

• To reproduce the results in Table 1 and 2 (test accuracy and running time of different backbone graph learning methods without unlearning):

./scripts/Exp1.sh

You can change the models or datasets that you want to test with by modifying Exp1.sh.

• To reproduce Figure 3 (test accuracy and running time of different unlearning methods, 10% sequential unlearning requests):

./scripts/Exp2.sh

• To reproduce Figure 4 (test accuracy and running time of different unlearning methods, 90% sequential unlearning requests):

./scripts/Exp3.sh

Please contact Chao Pan ([email protected]), Eli Chien ([email protected]) if you have any question.

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

@inproceedings{
pan2023unlearning,
title={Unlearning Graph Classifiers with Limited Data Resources},
author={Chao Pan and Eli Chien and Olgica Milenkovic},
booktitle={The Web Conference},
year={2023}
}