Certified Edge Unlearning for Graph Neural Networks

This is an implementation of paper [Certified Edge Unlearning for Graph Neural Networks](). We provide the code of graph edge unlearning. In addition, we present the code for reproducing the experiments.

Dependencies

• tqdm/numpy/pandas/sklearn/matplotlib/seaborn
• Pytorch >= 1.09
• pyg (torch_geometric)
• DeepRobust
• StellarGraph

Result Reproduction

Tightness of Bounds

python experiment.py -rq bound

Accuracy of CEU

python experiment.py -rq unlearn

Efficency of CEU

python experiment.py -rq efficiency

Efficacy of CEU

python experiment.py -rq efficacy

Effect of $\epsilon$

python experiment.py --rq epsilon

CGU Comparison

python experiment.py --rq cgu_compare

Common Parameters

• -g, the ID of a GPU you want to use. Default: -1 (using CPU)
• -edges, a list, indicates the numbers of edges you want to unlearn. Default: [100, 200, 400, 800, 1000].
• -targets, a list, indicates what target models you want to evaluate. Default:['gcn', 'sage', 'gin'].
• -datasets, a list, indicate what datasets you want to use. Default:['cora', 'citeseer', 'cs'].