Research map / Graph learning / GNN architectures
Gradient Rewiring for Editable Graph Neural Network Training
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 GRE/GRE+, a gradient-rewiring method that projects the target-node loss gradient away from a stored training-loss gradient so editing a GNN fixes a wrong node without degrading training-node loss.
Problem
Model editing for graph neural networks: correcting a single node's wrong prediction without disturbing predictions on other (training and unrelated) nodes, since GNN message passing spreads edit effects across the graph.
Approach
Observes large gradient inconsistency between the target-node loss and training-set loss in GNNs; stores the training-loss gradient as an anchor and rewires the target gradient by removing its component that conflicts with the anchor (closed-form QP in the dual), with GRE+ extending this to K disjoint training subsets.
Key results
- On A-photo with GCN (single edit), GRE+ reduces drawdown to 43.87% vs 65.08% for GD, with accuracy 50.13%.
- With EGNN-GCN on Cora, GRE+ achieves 88.99% accuracy with 0.05% drawdown vs 87.58%/1.42% for GD (single edit).
- GRE+ takes roughly 1.5-2.5x the wall-clock editing time of GD/GRE due to the QP solve.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GRE+ instead of GD | GCN · Cora, single edit | test drawdown: 5.03±6.40 -> 0.56±0.35 | Helped |
| Use GRE+ instead of GD | GCN · A-photo, single edit | test drawdown: 65.08±20.13 -> 43.87±1.36 | Helped |
| Use GRE+ instead of GD | GCN · ogbn-arxiv, single edit | test drawdown: 46.50±14.90 -> 15.39±1.10 | Helped |
| Use GRE+ instead of GD | GCN · Flickr, single edit | test drawdown: 37.25±10.20 -> 28.1±0.67 | Mixed |
| Add gradient rewiring to EGNN-GCN | EGNN-GCN · Coauthor-CS, single edit | test drawdown: 0.01±0.02 -> -0.32±0.06 | Helped |
| Add gradient rewiring to EGNN-GCN | EGNN-GCN · Reddit, single edit | test drawdown: 1.28±0.55 -> -0.21±0.38 | Helped |
| Replace GRE's single whole-training-set constraint with constraints on multiple disjoint training subsets (ablation) | GRE · A-photo, single edit | test drawdown: 58.12±2.26 -> 43.87±1.36 | Helped |
| Replace GRE's single whole-training-set constraint with constraints on multiple disjoint training subsets (ablation) | GRE · Flickr, single edit | test drawdown: 33.64±1.50 -> 28.1±0.67 | Helped |
| Use GRE+ rather than GRE (ablation) | GCN · sequential editing, A-photo | test drawdown | No clear effect |
| Increase the GRE+ subset count to 5 rather than use 2 or 3 subsets (ablation) | GRE+ · independent editing; datasets shown in the hyperparameter study | test drawdown | Hurt |
| Increase the GRE+ subset count from 2 to 5 (ablation) | GCN · ogbn-products, editing cost | editing time: 2413.69 -> 2591.06 | Hurt |
Surprising findings
- Editing with the target loss lowers target loss but raises training loss significantly, especially for GCN and GraphSAGE, while MLP degrades much less.
- ENN performs markedly worse than plain gradient descent on GNNs.
Limitations
- Effectiveness depends on the accuracy of the pre-stored training gradient, which may drift over long editing sequences.
- Sequential editing drawdown remains high (e.g., around 65% for GCN on A-photo), and GRE+ gives limited gains over GRE in that setting.
Experimental setup
Datasets: Cora A-computers A-photo Coauthor-CS Flickr Reddit ogbn-arxiv ogbn-products
Benchmarks: Cora A-computers A-photo Coauthor-CS Flickr Reddit ogbn-arxiv ogbn-products
Metrics: accuracy drawdown success rate edit time peak memory
Baselines: GD ENN EGNN
Builds on: EGNN Editable Neural Networks (ENN) gradient projection (GEM-style dual QP)
Closest papers by meaning
- Locality-Aware Graph Rewiring in GNNs (ICLR 2024 · similarity 0.88)
- GNNs Also Deserve Editing, and They Need It More Than Once (ICML 2024 · similarity 0.88)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.87)
- Erase Then Rectify: A Training-Free Parameter Editing Approach for Cost-Effective Graph Unlearning (AAAI 2025 · similarity 0.86)
- GNNs Getting ComFy: Community and Feature Similarity Guided Rewiring (ICLR 2025 · similarity 0.86)
- AdapterGNN: Parameter-Efficient Fine-Tuning Improves Generalization in GNNs (AAAI 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.