Research map / Graph learning / GNN architectures
IntraMix: Intra-Class Mixup Generation for Accurate Labels and Neighbors
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 IntraMix, which applies Mixup only among same-pseudo-label nodes to create higher-quality labeled nodes and connects each generated node to high-confidence same-class nodes to enrich neighborhoods, as a plug-in for GNNs.
Problem
GNNs on graphs suffer from insufficient accurate labels and limited neighbors per node; vanilla Mixup performs poorly for node classification because generated nodes mix classes and have no clear neighbors.
Approach
Perform Mixup only between nodes sharing the same (pseudo-)label so generated nodes have a single class, then attach each generated node to two high-confidence same-class nodes chosen by consistency across dropout-perturbed GNN predictions, with theory showing reduced label and propagated noise.
Key results
- On Cora with GCN, IntraMix reaches 85.25% vs 81.51% original and 83.47% for the best prior baseline (NodeMixup/CODA), in semi-supervised node classification averaged over 30 runs.
- IntraMix improves GCN on ogbn-arxiv 1% split from 62.99% to 64.84% and on Flickr 1% from 46.02% to 48.18%.
- Ablation on GCN: Intra-Class Mixup gives 85.2 on Cora vs 83.0 for vanilla Mixup with connection to source nodes and 58.9 without connection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Intra-Class Mixup generation (vs. pseudo-labeling only) (ablation) | GCN · Cora, CiteSeer, Pubmed | accuracy (%): Cora 82.9 to 85.2; CiteSeer 72.3 to 74.8; Pubmed 79.5 to 82.9 | Helped |
| Vanilla Mixup with generated nodes isolated (no connections) (ablation) | GCN · Cora | accuracy (%): 81.5 original to 58.9 with high variance | Hurt |
| Vanilla Mixup with generated nodes connected to their source nodes (ablation) | GCN · Cora | accuracy (%): 83.0 vs 82.9 for pseudo-labeling only; well below IntraMix 85.2 | No clear effect |
| Replacing generated nodes with all-one vectors (ablation) | GCN · Cora | accuracy (%): 81.5 original to 31.9 | Hurt |
| Replacing generated nodes with all-zero vectors (neighbor connections kept) (ablation) | GCN · Cora | accuracy (%): 81.5 original to 83.8, below IntraMix 85.2 | Helped |
| Neighbor selection: connect generated nodes to random nodes (ablation) | GCN · Cora | accuracy (%): 81.5 original to 76.7 | Hurt |
| Neighbor selection: connect generated nodes to high-confidence same-class nodes (IntraMix) vs. direct connection to low-quality same-class nodes (ablation) | GCN · Cora | accuracy (%): Direct Con 83.6 vs IntraMix 85.2 | Helped |
| IntraMix as plug-in augmentation across GNNs (GCN, GAT, SAGE, APPNP) | GCN, GAT, SAGE, APPNP · Cora, CiteSeer, Pubmed, CS, Physics (semi-supervised) | accuracy (%): Best or near-best on most dataset/model pairs versus Original and baselines | Helped |
| IntraMix on large-scale graphs (ogbn-arxiv, Flickr) with 1% and 5% labels and full supervision | GCN, GAT, SAGE · ogbn-arxiv, Flickr | accuracy (%): e.g., ogbn-arxiv 1% GCN 62.99 to 64.84; Flickr 1% GCN 46.02 to 48.18 | Helped |
| IntraMix in inductive setting | GAT, SAGE · Cora, CiteSeer (inductive) | accuracy (%): GAT Cora 81.3 to 83.8; SAGE CiteSeer 69.1 to 73.9 | Helped |
| IntraMix on heterogeneous graphs (SeHGNN base) | SeHGNN · IMDB, DBLP | accuracy: qualitative, values in figure | Helped |
Surprising findings
- Vanilla Mixup without connections collapses performance (58.9 on Cora with large variance), and zero-vector replacement of generated nodes still improves over the original GCN, suggesting neighbor selection alone contributes gains.
- Mixup-based augmentation is claimed to be limited for node classification in prior work, but IntraMix shows gains when Mixup is restricted to same-label nodes.
Limitations
- Neighbor selection relies on the homophily/neighborhood assumption, and gains on heterophilic graphs are shown only in a figure.
- Generated node count is manually set and there is no principled way to choose it.
- Theoretical guarantees rest on simplifying assumptions (Gaussian label noise, f with additive property, linearized GNN).
Experimental setup
Datasets: Cora CiteSeer Pubmed CS Physics ogbn-arxiv Flickr Cornell Texas Wisconsin IMDB DBLP
Benchmarks: Cora CiteSeer Pubmed CS Physics ogbn-arxiv Flickr Cornell Texas Wisconsin IMDB DBLP
Metrics: accuracy MADgap
Baselines: GraphMix CODA FLAG DropMessage MH-Aug GRAND Local Augmentation NodeMixup LA-GCN UPS
Builds on: Mixup GraphMix UPS GRAND NodeMixup
Closest papers by meaning
- AGMixup: Adaptive Graph Mixup for Semi-supervised Node Classification (AAAI 2025 · similarity 0.90)
- NodeMixup: Tackling Under-Reaching for Graph Neural Networks (AAAI 2024 · similarity 0.89)
- Graph Mixup on Approximate Gromov–Wasserstein Geodesics (ICML 2024 · similarity 0.86)
- Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization (AAAI 2024 · similarity 0.86)
- NLGT: Neighborhood-based and Label-enhanced Graph Transformer Framework for Node Classification (AAAI 2025 · similarity 0.85)
- Diffusing to the Top: Boost Graph Neural Networks with Minimal Hyperparameter Tuning (ICLR 2025 · similarity 0.85)
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.