Research map / Graph learning / GNN architectures
NodeMixup: Tackling Under-Reaching for Graph Neural Networks
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.
Identifies under-reaching in GNN semi-supervised node classification and proposes NodeMixup, an architecture-agnostic labeled-unlabeled mixup with neighbor-label-distribution-aware sampling to improve reachability.
Problem
Labeled nodes are sparsely distributed, so low-degree unlabeled nodes far from labels cannot receive supervision within shallow GNNs (under-reaching), causing poor distribution alignment between labeled and unlabeled nodes.
Approach
Mix labeled nodes with pseudo-labeled unlabeled nodes (cross-set pairing), fuse neighbor adjacency for intra-class pairs while blocking message passing for inter-class pairs with an identity matrix, and sample mixing partners by neighbor-label-distribution similarity and low degree.
Key results
- On Cora with GCN, NodeMixup reaches 83.47% vs 81.57% for the original GCN, and on Physics it reaches 93.97% vs 92.56%.
- On ogbn-arxiv, GCN+NodeMixup achieves 73.46% test accuracy vs 71.74% for GCN, surpassing GCNII (72.74%).
- Ablation on Cora with GCN: LU pairing yields 83.64% vs 81.67% for vanilla mixup and 82.41% for random sampling instead of NLD-aware sampling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add NodeMixup | GCN · Cora | accuracy: 81.57 -> 83.47 | Helped |
| Add NodeMixup | GAT · Citeseer | accuracy: 71.82 -> 74.30 | Helped |
| Add NodeMixup | APPNP · CS | accuracy: 91.79 -> 92.82 | Helped |
| Add NodeMixup | GraphSAGE · Physics | accuracy: 93.09 -> 94.76 | Helped |
| Use labeled-unlabeled rather than labeled-labeled pairing (ablation) | GCN + NodeMixup · Cora | accuracy: 81.85 -> 83.64 | Helped |
| Use labeled-unlabeled rather than labeled-all pairing (ablation) | GCN + NodeMixup · Cora | accuracy: 81.97 -> 83.64 | Helped |
| Add inter-class mixup to intra-class mixup (ablation) | GCN + NodeMixup · Cora | accuracy: 82.06 -> 83.64 | Helped |
| Add intra-class mixup to inter-class mixup (ablation) | GCN + NodeMixup · Cora | accuracy: 81.97 -> 83.64 | Helped |
| Replace random sampling with NLD-aware sampling (ablation) | GCN + NodeMixup · Cora | accuracy: 82.41 -> 83.64 | Helped |
| Use vanilla mixup (ablation) | GCN · Cora | accuracy: 81.57 -> 81.67 | No clear effect |
| Use intra-class mixup only (ablation) | GCN + NodeMixup · Cora | accuracy | No clear effect |
Surprising findings
- Intra-class-only interpolation does not significantly improve the backbone model, according to the ablation (IR mixup 82.06% on Cora).
Limitations
- Hyperparameters (lambda, beta, gamma) are grid-searched per backbone, and baseline methods use their original settings rather than being re-tuned.
- Ablations are reported only on Cora with GCN.
- Under-reaching analysis is primarily empirical on Cora; the theoretical grounding is limited.
Experimental setup
Datasets: Cora Citeseer Pubmed Coauthor CS Coauthor Physics ogbn-arxiv
Benchmarks: Cora Citeseer Pubmed Coauthor CS Coauthor Physics ogbn-arxiv
Metrics: classification accuracy (%)
Baselines: GCN GAT APPNP GraphSAGE PASTEL ReNode GraphMix UPS JKNet GCNII
Builds on: mixup GraphMix neighbor label distribution reaching coefficient centered kernel alignment
Closest papers by meaning
- IntraMix: Intra-Class Mixup Generation for Accurate Labels and Neighbors (NeurIPS 2024 · similarity 0.89)
- AGMixup: Adaptive Graph Mixup for Semi-supervised Node Classification (AAAI 2025 · similarity 0.88)
- Diffusing to the Top: Boost Graph Neural Networks with Minimal Hyperparameter Tuning (ICLR 2025 · similarity 0.88)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.86)
- Normalize Then Propagate: Efficient Homophilous Regularization for Few-Shot Semi-Supervised Node Classification (AAAI 2025 · similarity 0.86)
- Improving Graph Neural Networks by Learning Continuous Edge Directions (ICLR 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.