Graph unlearning based on optimized saliency parameters

Yu Yang

doi:10.1016/j.hcc.2026.100446

Graph unlearning based on optimized saliency parameters

2026-10-21

0

OA

AI

DOI：10.1016/j.hcc.2026.100446

Original

Save

Abstract

En 中文

The need to remove specific information from trained graph neural networks (GNNs) has grown increasingly important in recent years. However, existing solutions such as full retraining or indiscriminate parameter adjustment are computationally expensive and often result in reduced model performance or incomplete unlearning. To address these limitations, we propose a novel graph unlearning method based on optimized saliency parameters. This approach achieves effective unlearning while maintaining overall utility of the model. In this proposed method, we first analyze node prediction errors through multiple forward passes to capture their global statistical patterns and identify representative nodes that have the greatest impact on model performance. Second, we estimate the importance of samples to be unlearned and retained using the Fisher information matrix, and construct parameter-level masks to guide selective updates during the process of unlearning. Finally, we formulate a dual-objective optimization problem that jointly considers unlearning effectiveness and retention performance, and determine the optimal balance between them through Pareto front analysis. Extensive experiments on multiple GNNs and real-world datasets demonstrate that our proposal consistently outperforms existing approaches in terms of unlearning quality, model performance, and computational efficiency.

Keywords:

Graph unlearning

Saliency parameters

Graph neural networks

Model utility

Graph unlearning based on optimized saliency parameters

Abstract

Journal

Organization

Cited Papers