Research map / Graph learning / GNN architectures
NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise
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 NoisyGL, a unified benchmark library evaluating 7 learning-with-label-noise and 10 graph-learning-under-label-noise methods on 8 node classification datasets under uniform, pair, and asymmetric noise.
Problem
Node classification with GNNs degrades under label noise, but existing graph-label-noise (GLN) studies use inconsistent datasets, splits, noise settings and preprocessing, so methods cannot be fairly compared.
Approach
A unified open-source benchmark with standardized backbones, APIs, data splits and label-corruption procedure, used to compare LLN and GLN methods and analyze how graph structure spreads label noise (e.g., via metrics AUCS/AUU/AUIS and misleading train accuracy).
Key results
- Most LLN methods (e.g., S-model, Forward, Backward, Coteaching, JoCoR) perform statistically similar to GCN under noise; APL and SCE give slight gains.
- No GLN method consistently beats GCN across all datasets; on Flickr none of the GLN methods outperform the GCN baseline.
- RNCGLN is the slowest, taking 66.8x GCN's time on Cora and 2945.8x on DBLP, and runs out of memory on PubMed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply traditional LLN methods to GCN | GCN · Graph node classification under label noise across datasets | test accuracy | No clear effect |
| Apply APL or SCE robust loss | GCN · Graph node classification under label noise across datasets | test accuracy | Helped |
| Use GLN methods instead of the GCN baseline | GCN · Node classification under label noise across datasets | test accuracy | Mixed |
| Use NRGNN instead of GCN | GCN · Cora, Citeseer, and DBLP under label noise | test accuracy | Helped |
| Use GLN methods instead of GCN | GCN · Flickr under label noise | test accuracy | Hurt |
| Increase label noise rate, especially with pair noise | Graph classification models · DBLP | test accuracy | Hurt |
| Use pair noise instead of uniform noise | GCN · Cora, Citeseer, Pubmed, Amazon-Computers, and BlogCatalog at 50% noise | misleading train accuracy | Hurt |
| Add graph structure augmentation | GLN methods · Cora and Amazon-Photos under 30% uniform noise | AUIS: 51.55 ± 6.53 -> 78.81 ± 5.94 | Helped |
| Incorrectly supervise neighboring nodes rather than correctly supervise them | GCN · Cora under 30% uniform noise | unlabeled-node accuracy | Hurt |
| Use a sparse graph rather than a dense graph | GCN · Graph node classification under label noise | spread effect of label noise | Hurt |
Surprising findings
- On Flickr, none of the GLN methods outperform the plain GCN baseline.
- Pair noise is more harmful than uniform noise; at 50% pair noise test accuracy of all methods almost halves.
Limitations
- Datasets are predominantly homophilous; GLN methods struggle on heterophilous graphs like Flickr.
- Only uniform and pair (plus asymmetric in appendix) instance-independent noise considered; instance-dependent noise not studied.
- Limited set of implemented methods; not covering graph classification or link prediction.
Experimental setup
Datasets: Cora Citeseer Pubmed DBLP Amazon-Computers Amazon-Photo BlogCatalog Flickr
Benchmarks: Cora Citeseer Pubmed DBLP Amazon-Computers Amazon-Photo BlogCatalog Flickr
Metrics: test accuracy standard deviation over 10 runs runtime ACLT AILT AUCS AUU AUIS misleading train accuracy
Baselines: GCN NRGNN RTGNN CP D-GNN RNCGLN CLNode PIGNN UnionNET CGNN CRGNN Coteaching
Builds on: GCN NRGNN RTGNN CP CLNode
Closest papers by meaning
- A Simple and Yet Fairly Effective Defense for Graph Neural Networks (AAAI 2024 · similarity 0.89)
- Training Robust Graph Neural Networks by Modeling Noise Dependencies (NeurIPS 2025 · similarity 0.89)
- Label-free Node Classification on Graphs with Large Language Models (LLMs) (ICLR 2024 · similarity 0.88)
- IGL-Bench: Establishing the Comprehensive Benchmark for Imbalanced Graph Learning (ICLR 2025 · similarity 0.88)
- Mitigating Label Noise on Graphs via Topological Sample Selection (ICML 2024 · similarity 0.88)
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.88)
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.