Research map / Graph learning / GNN architectures
Training Robust Graph Neural Networks by Modeling Noise Dependencies
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.
Introduces dependency-aware noise on graphs (DANG), where feature noise propagates to structure and labels, and proposes DA-GNN, a variational deep generative robust GNN with new Auto/Garden DANG benchmarks.
Problem
Robust GNN training under node feature noise that propagates to graph structure and node labels, a setting not covered by existing methods that assume at least one clean data source (independent feature noise assumption).
Approach
Models the data generating process of DANG (X causes A and Y; epsilon causes X and A; Z_Y and Z_A latent clean labels and structure) with a variational ELBO, instantiated via GCN-based encoders/decoders, edge reconstruction with label smoothing, a homophily regularizer, and small-loss estimation of clean edges.
Key results
- On synthetic DANG-50% Cora, DA-GNN reaches 69.7% accuracy vs 66.4% for the next best (ProGNN).
- On real-world Auto link prediction under DANG, DA-GNN gets 73.6 ROC-AUC vs 74.8 for GraphGLOW (best), i.e. DA-GNN is not best there.
- On Garden node classification under DANG, DA-GNN ties SG-GSR at 80.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| model the DANG causal relationships with DA-GNN rather than use the baselines | node classification on synthetic DANG · Cora, Citeseer, Photo, Computers, and Arxiv, DANG-10%, DANG-30%, and DANG-50% | accuracy | Helped |
| use DA-GNN rather than the strongest baseline | node classification on real-world DANG · Auto + DANG | accuracy: 62.0±1.1 vs 61.4±0.4 | No clear effect |
| use DA-GNN rather than the strongest baseline | link prediction on real-world DANG · Auto + DANG | ROC-AUC: 74.8±0.2 vs 73.6±0.6 | Hurt |
| use DA-GNN rather than the strongest baseline | node classification on real-world DANG · Garden + DANG | accuracy: 80.2±0.4 vs 80.2±0.8 | No clear effect |
| use DA-GNN rather than the strongest baseline | link prediction on real-world DANG · Garden + DANG | ROC-AUC: 91.2±0.5 -> 92.4±0.4 | Helped |
| remove the Y ← (X,A) causal relationship (ablation) | DA-GNN · DANG node classification | accuracy | Hurt |
| remove the A ← X causal relationship (ablation) | DA-GNN · DANG node classification | accuracy | Hurt |
| remove the A ← ε causal relationship (ablation) | DA-GNN · DANG node classification | accuracy | Hurt |
Surprising findings
- Under DANG on Auto link prediction, GraphGLOW (74.8) outperforms DA-GNN (73.6), contrary to the claim of consistent superiority.
Limitations
- Does not model X <- A (structure noise causing feature noise), so not all noise scenarios covered.
- Relies on assumption that label noise arises from features and structure rather than external factors (epsilon not a cause of Y).
- Some baselines (e.g., on Arxiv) run out of memory, and DA-GNN has hyperparameters like k requiring tuning.
Experimental setup
Datasets: Cora Citeseer Amazon Photo Amazon Computers ogbn-arxiv Amazon Auto Amazon Garden PolitiFact
Benchmarks: Cora Citeseer Photo Computers Arxiv Auto Garden
Metrics: accuracy ROC-AUC
Baselines: WSGNN GraphGLOW AirGNN ProGNN RSGNN STABLE EvenNet NRGNN RTGNN SG-GSR
Builds on: variational autoencoder CausalNL small-loss approach GCN
Closest papers by meaning
- A Simple and Yet Fairly Effective Defense for Graph Neural Networks (AAAI 2024 · similarity 0.89)
- NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise (NeurIPS 2024 · similarity 0.89)
- Learn Beneficial Noise as Graph Augmentation (ICML 2025 · similarity 0.88)
- Accurate and Scalable Estimation of Epistemic Uncertainty for Graph Neural Networks (ICLR 2024 · similarity 0.88)
- NoiseHGNN: Synthesized Similarity Graph-Based Neural Network for Noised Heterogeneous Graph Representation Learning (AAAI 2025 · similarity 0.87)
- Bounding the Expected Robustness of Graph Neural Networks Subject to Node Feature Attacks (ICLR 2024 · similarity 0.87)
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.