Research map / Graph learning / Graph self-supervised learning
Robust Node Classification on Graph Data with Graph and 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.
Proposes RNCGLN, a self-attention-based node classifier that jointly handles graph noise and label noise via a graph contrastive loss, pseudo graphs, and pseudo labels.
Problem
Node classification on graph data where graph structure (missing/noisy edges) and labels are both noisy, a setting prior robust methods handle only one of at a time.
Approach
Uses multi-head self-attention as backbone for global graph learning, a multi-hop graph contrastive loss for local graph learning, and two self-improvement modules that iteratively replace the graph with a pseudo graph and the labels with confident pseudo labels.
Key results
- RNCGLN improves average accuracy by 3.50% over the best comparison method (LAFAK) across four datasets.
- Under {gr, lr}={0.1,0.3} on Cora, RNCGLN reaches 81.73±0.33 vs 79.34±0.54 for LAFAK.
- Ablation: removing the graph contrastive loss drops performance by 19.95% on average under noisy settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RNCGLN instead of the strongest baseline | Node classification · Cora, Citeseer, Pubmed, Photo; clean and noisy settings in Table 1 | accuracy: 81.73±0.33 vs 79.34±0.54; 79.23±1.03 vs 75.61±1.54; 90.72±0.08 vs 85.83±0.34 | Helped |
| Add graph contrastive loss for local graph learning (ablation) | RNCGLN · Cora {gr, lr}={.1, .3} | accuracy: 53.93±0.66 -> 81.73±1.42 | Helped |
| Add node selection (ablation) | RNCGLN · Cora {gr, lr}={.1, .3} | accuracy: 81.56±1.56 -> 81.73±1.42 | Helped |
| Add pseudo graph and pseudo label self-improvement (ablation) | RNCGLN · Cora, Citeseer, Pubmed, Photo under both noisy settings | accuracy: 75.90±1.01 -> 81.73±1.42 | Helped |
| Add pseudo label while retaining pseudo graph (ablation) | RNCGLN · Cora {gr, lr}={.1, .3} | accuracy: 78.36±0.60 -> 81.73±1.42 | Helped |
| Add pseudo graph while retaining pseudo label (ablation) | RNCGLN · Cora {gr, lr}={.1, .3} | accuracy: 80.80±0.90 -> 81.73±1.42 | Helped |
| Increase label-noise ratio | RNCGLN · Cora, Citeseer, Pubmed, Photo | accuracy | Hurt |
| Increase graph-noise ratio | RNCGLN · Cora, Citeseer, Pubmed, Photo | accuracy | Hurt |
| Vary loss weight α (ablation) | RNCGLN · Cora and Citeseer; {gr, lr}={0.3, 0.1} and {0.1, 0.3} | accuracy | Mixed |
Surprising findings
- Graph-noise robust GNNs such as IDGL can perform worse than plain GCN under label noise (e.g., on Photo under {0.1,0.3}).
Limitations
- Relies on hyper-parameters (alpha, thresholds tau_p1, tau_p2, tau_g1, tau_g2) that require tuning.
- Evaluated only on four datasets with synthetic noise injection.
Experimental setup
Datasets: Cora Citeseer Pubmed Photo
Benchmarks: Cora Citeseer Pubmed Photo
Metrics: ACC standard deviation
Baselines: GCN GAT JLGCN IDGL NRGNN LAFAK
Builds on: Transformer self-attention efficient attention graph contrastive loss (Hu et al. 2021) self-training pseudo labels
Closest papers by meaning
- Graph Attention is Not Always Beneficial: A Theoretical Analysis of Graph Attention Mechanisms via Contextual Stochastic Block Models (ICML 2025 · similarity 0.87)
- Joint Graph Rewiring and Feature Denoising via Spectral Resonance (ICLR 2025 · similarity 0.87)
- Mitigating Label Noise on Graphs via Topological Sample Selection (ICML 2024 · similarity 0.87)
- NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise (NeurIPS 2024 · similarity 0.87)
- Training Robust Graph Neural Networks by Modeling Noise Dependencies (NeurIPS 2025 · similarity 0.87)
- Noisy Node Classification by Bi-level Optimization Based Multi-Teacher Distillation (AAAI 2025 · 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.