A PyTorch implementation of "Graph Unlearning meets Influence-aware Negative Preference Optimization".

INPO is a PO-based graph unlearning method to improving the robustness of the model utility to the unlearning process. The core idea behind INPO is to amplify the influence of unlearned edges and mitigate the tight topological coupling between the forget set and the retain set, reducing impact on model utility when unlearning specific edges.

We employ standard and commonly used datasets, which you can download directly according to given links.

To reproduce our results, please run:

bash run.sh

For easy reproduction, we provide detailed hyper-parameter setting. Implementation of INPO-S based on the original source code(https://github.com/mims-harvard/GNNDelete) , and we provide file INPO-S.py in our code.