Research map / Graph learning / GNN architectures
A Cognac Shot To Forget Bad Memories: Corrective Unlearning for Graph Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoGNAC, a corrective unlearning method for GNNs that alternates contrastive unlearning on affected graph neighborhoods with ascent-descent unlearning, recovering most oracle accuracy after manipulations are only partly identified.
Problem
Removing the adverse effects of manipulated (label-flipped, spurious-edge, or trigger-poisoned) training entities from a trained GNN when only a small representative subset of the manipulated set is identified, since message passing propagates their influence to neighbors.
Approach
Combines a graph-aware contrastive step that identifies affected neighbors via feature-inversion logit change and pushes them away from the deletion set toward unaffected neighbors, with an ascent-descent step (gradient ascent on the deletion set, descent on the retain set) applied to the GCN.
Key results
- On CS and Cora, CoGNAC reaches 79.3 and 75.5 affected accuracy versus 44.3 and 51.4 for the original model and 90.1 and 71.1 for the Oracle (CS and Cora ablation).
- CoGNAC is reported to be 8x more efficient than MEGU's sampling and to outperform retraining from scratch on most settings, unlearning with only 5% of the manipulation set identified.
- On PubMed with 38.96% of training nodes in S_f, CoGNAC retains strong performance, as stated in the paper's large-S_f experiments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CoGN to ACDC (ablation) | Cognac · CS and Cora | affected accuracy: 67.2 -> 79.3; 59.9 -> 75.5 | Helped |
| Add ACDC to CoGN (ablation) | Cognac · CS and Cora | affected accuracy: 47.2 -> 79.3; 42.1 -> 75.5 | Helped |
| Use Cognac rather than retraining from scratch with partial manipulation knowledge | GCN corrective unlearning · Affected distributions across datasets and manipulations | affected accuracy | Helped |
| Use Cognac rather than existing graph unlearning methods | GCN corrective unlearning · Targeted edge and label attacks | affected accuracy | Helped |
| Add CoGN to ACDC (ablation) | Cognac · CS and Cora | remaining accuracy: 86.6 -> 82.3; 61.3 -> 56.6 | Hurt |
| Replace the affected-node identification heuristic with MEGU's sampling technique (ablation) | Cognac affected-node identification · Affected-node identification evaluation | performance: over 25% | Hurt |
| Use Cognac rather than SCRUB | GCN corrective unlearning · DBLP (29.4% deletion) | affected accuracy: 81.7 vs 82.1 | Hurt |
Surprising findings
- Existing graph unlearning methods (GNNDelete, GIF, MEGU, UtU) fail to improve affected accuracy even when the whole manipulation set is known, while naive finetuning and SCRUB-like methods without graph components do better.
- Retraining from scratch on the remaining data can reinforce the manipulation, so it is not a gold standard for corrective unlearning.
Limitations
- Relies on the homophily assumption and is evaluated mainly on homophilic datasets, so heterophilic graphs are not covered.
- Does not guarantee successful unlearning against arbitrary real-world attacks.
- Theoretical guarantees rest on strong simplifying assumptions (convexity, bounded gradients) and a first-order approximation.
Experimental setup
Datasets: Cora PubMed DBLP Coauthor CS Coauthor Physics Amazon Photos Amazon Computers CiteSeer CoraFull OGB-arXiv
Benchmarks: Cora CiteSeer PubMed DBLP Coauthor CS Coauthor Physics Amazon Photos Amazon Computers CoraFull OGB-arXiv
Metrics: affected-distribution accuracy (Acc_aff) remaining-entity accuracy (Acc_rem) unlearning time
Baselines: GNNDelete GIF MEGU UtU SCRUB Finetune Retrain Oracle
Builds on: GNNDelete MEGU SCRUB Corrective unlearning (Goel et al.) GCN
Closest papers by meaning
- Erase Then Rectify: A Training-Free Parameter Editing Approach for Cost-Effective Graph Unlearning (AAAI 2025 · similarity 0.90)
- Towards Effective and General Graph Unlearning via Mutual Evolution (AAAI 2024 · similarity 0.88)
- Community-Centric Graph Unlearning (AAAI 2025 · similarity 0.88)
- Efficient Unlearning for Spatio-temporal Graph (Student Abstract) (AAAI 2025 · similarity 0.87)
- Not All Wrong is Bad: Using Adversarial Examples for Unlearning (ICML 2025 · similarity 0.87)
- Locality-Aware Graph Rewiring in GNNs (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.