Contrastive unlearning via representation editing for graph neural networks

Graph unlearning (GU) aims to eliminate the influence of specific nodes, edges, or features from a trained graph neural network (GNN), and is of great importance for privacy protection and data quality management. However, existing GU methods often struggle to achieve a satisfactory balance between model utility and unlearning effectiveness, and they also find it difficult to remove the influence of target data and its propagated effects over the graph. To address these issues, we propose a GU framework named CURE based on contrastive representation editing. CURE introduces an adaptive sample selection module to accurately identify key nodes that are highly related to the unlearning targets from both structural and semantic perspectives. It then adopts a contrastive unlearning strategy to decouple the representations of the forgotten nodes from those of their related nodes in the embedding space, thereby effectively removing the information of the forgotten nodes. In addition, to preserve the predictive capacity of the model on the remaining data, we further introduce a personalized PageRank-based stability preservation module, which constrains the consistency of prediction distributions for affected nodes before and after unlearning. The experimental results show that CURE achieves a favorable trade-off among model utility, unlearning efficiency, and unlearning effectiveness, and outperforms existing baseline methods in most experimental settings, while also exhibiting superior privacy protection and robustness.