Certified Signed Graph Unlearning
Junpeng Zhao, Lin Li, Kaixi Hu, Kaize Shi, Jingling Yuan
TL;DR
This paper introduces Certified Signed Graph Unlearning (CSGU), a method that ensures privacy and maintains utility in signed graph neural networks by identifying influenced neighborhoods, applying sociological importance measures, and importance-weighted updates.
Contribution
The paper presents the first unlearning method tailored for signed graph neural networks, incorporating sociological theories for improved privacy and utility guarantees.
Findings
CSGU outperforms existing unlearning methods on signed graph neural networks.
It provides provable privacy guarantees while maintaining high model utility.
Experiments show significant improvements in unlearning effectiveness and utility preservation.
Abstract
Signed graphs model complex relationships through positive and negative edges, with widespread real-world applications. Given the sensitive nature of such data, selective removal mechanisms have become essential for privacy protection. While graph unlearning enables the removal of specific data influences from Graph Neural Networks (GNNs), existing methods are designed for conventional GNNs and overlook the unique heterogeneous properties of signed graphs. When applied to Signed Graph Neural Networks (SGNNs), these methods lose critical sign information, degrading both model utility and unlearning effectiveness. To address these challenges, we propose Certified Signed Graph Unlearning (CSGU), which provides provable privacy guarantees while preserving the sociological principles underlying SGNNs. CSGU employs a three-stage method: (1) efficiently identifying minimal influenced…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
S1. The paper addresses an important and underexplored problem—unlearning in signed graphs—which is both timely and meaningful. S2. The proposed components—Triadic Influence Neighborhood and Sociological Influence Quantification—are conceptually intuitive and well-motivated.
W1. My primary concern is that CSGU does not truly achieve certified unlearning, as the introduction of edge weights fundamentally modifies the original loss function. Specifically, CSGU guarantees similarity between the unlearned model and the retrained model with respect to the weighted loss defined in Eq. (7), rather than the original unweighted loss. W2. The computation of triadic closure is computationally expensive and poses a significant scalability bottleneck for real-world graph applic
S1 Originality: The paper tackles machine unlearning in the signed graph domain, a niche yet practically important area where relationships can be both positive and negative. S2 Quality: The theoretical analysis is rigorous, with formal derivations of the certification bound and proofs ensuring robustness of the unlearning process. S3 Clarity: The paper is clearly structured, progressing logically from problem motivation to theoretical formulation, algorithm, and experiments. Figures (particul
W1 While the method is well-motivated for signed GNNs, its broader applicability to general (unsigned) GNNs or heterogeneous graphs is not fully demonstrated. More importantly, it would be helpful to discuss techniques that normalize the signed weights on graphs. W2 The certification bound relies on the assumption that node influence diminishes exponentially with hop distance. While this is practical, it might not hold for densely connected or long-range dependency graphs. W3 Although results
- This paper proposed the first certified unlearning method for SGNNs. Extending to signed graphs and defining suitable certificates under the signed Laplacian is technically nontrivial. - The authors proposed a novel upper bound on node influence of signed graphs and a weighting scheme that allocates the privacy budget based on the sociological importance of each edge. - Experiments are consistent across datasets, metrics, and removal settings. The ablations (e.g., the effect of truncation r
- The unlearning certification in the theoretical analysis seems to be standard. It would be helpful to highlight the challenges of certified unlearning on signed graphs. - Some important baselines are missing [1]. - The ablation of the privacy budget weighting seems to be missing in the ablation study. This appears to be one of the major contributions of the method, and it would be helpful to clarify this further. [1] Wu K, Shen J, Ning Y, et al. Certified edge unlearning for graph neural ne
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Privacy-Preserving Technologies in Data · Machine Learning in Healthcare