Research map / Graph learning / GNN architectures
GNNs Also Deserve Editing, and They Need It More Than Once
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 SEED-GNN, a batching-based editing method for graph neural networks that mitigates overfitting to achieve sequential editing robustness, built on the frozen GNN + active MLP design of EGNN.
Problem
Editing GNNs sequentially (multiple successive corrections on node classification) causes large test accuracy drawdown and overwritten earlier edits; existing GNN editors like EGNN fail beyond a few edits.
Approach
Attribute sequential editing failure to overfitting on edit targets and mitigate it by batching each edit with all previous edit targets, the target's training neighbors (neighborhood corrosion mitigation), and random training samples, while keeping the GNN frozen and updating only an active MLP.
Key results
- On ogbn-products with GCN, SEED-GNN keeps Avg SR at 1.0 with Avg DD of 8.3 across 50 sequential edits, while EGNN reaches Avg SR 0.72 and Avg DD 30.4.
- On Amazon Computers with GCN (50 edits), SEED-GNN achieves Avg SR 1.0 and Avg DD 1.1, versus EGNN's Avg SR 0.47 and Avg DD 22.6.
- Batching 100 random training samples with each edit target sharply reduces drawdown for vanilla GNNs (e.g., GCN on Amazon Computers 10th-edit DD drops from 32.9 to 8.0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SEED-GNN instead of EGNN | GCN · Amazon Computers | Avg DD: 22.6 -> 1.1 | Helped |
| use SEED-GNN instead of EGNN | GCN · ogbn-arxiv | Avg SR: 0.40 -> 1.0 | Helped |
| use SEED-GNN instead of EGNN | GCN · ogbn-products | Avg DD: 30.4 -> 8.3 | Helped |
| use SEED-GNN instead of EGNN | GCN · Cora | Avg SR: 0.60 -> 1.0 | Helped |
| use SEED-GNN instead of EGNN | GCN · Amazon Photo | Avg DD: 23.4 -> 0.8 | Helped |
| batch randomly selected training samples with the edit target (ablation) | vanilla GNN, GCN on Amazon Computers · Amazon Computers | Avg DD: 63.9 -> 4.7 | Helped |
| batch randomly selected training samples with the edit target (ablation) | Frozen GNN+MLP, GCN on Amazon Computers · Amazon Computers | Avg DD: 22.6 -> -0.6 | Helped |
| batch randomly selected training samples with the edit target (ablation) | vanilla GNN, GraphSAGE on ogbn-arxiv · ogbn-arxiv | Avg DD: 59.3 -> 7.8 | Helped |
| batch randomly selected training samples with the edit target (ablation) | Frozen GNN+MLP, GraphSAGE on ogbn-arxiv · ogbn-arxiv | Avg DD: 17.8 -> 1.4 | Helped |
| edit an instance with vanilla gradient descent rather than SEED-GNN batching (ablation) | GCN editing · Cora, 1-hop neighbors | accuracy: 73.4 -> 54.8 | Hurt |
Surprising findings
- Overfitting to the single edit target, rather than neighborhood aggregation alone, is argued to be the main cause of GNN edit sensitivity, contrary to EGNN's explanation.
- Standard single-edit fine-tuning or ENN can cause over 92.4% accuracy drop, making edited models unusable.
Limitations
- Limited to node classification; other graph tasks such as edge prediction and graph classification are not explored.
- Max drawdown remains high in some settings despite low average drawdown.
- Not recommended for high-stakes deployment without further evaluation.
Experimental setup
Datasets: Cora Amazon Photo Coauthor-CS Amazon Computers ogbn-arxiv ogbn-products YelpCHI
Benchmarks: Cora Amazon Photo Coauthor-CS Amazon Computers ogbn-arxiv ogbn-products YelpCHI
Metrics: i-th Edits Test Drawdown i-th Edits Success Rate Max Drawdown Average Drawdown Average Success Rate
Baselines: FT ENN EGNN Adapter-tuning LoRA
Builds on: EGNN MEND Adapter-tuning LoRA
Closest papers by meaning
- Gradient Rewiring for Editable Graph Neural Network Training (NeurIPS 2024 · similarity 0.88)
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.86)
- Stealth edits to large language models (NeurIPS 2024 · similarity 0.86)
- Perturbation-Restrained Sequential Model Editing (ICLR 2025 · similarity 0.85)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.85)
- Adversarial Representation Engineering: A General Model Editing Framework for Large Language Models (NeurIPS 2024 · similarity 0.85)
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.