Research map / Graph learning / Graph self-supervised learning
AGMixup: Adaptive Graph Mixup for Semi-supervised Node Classification
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 AGMixup, a subgraph-centric mixup for semi-supervised node classification whose mixing ratio is adaptively set from subgraph contextual similarity and predictive uncertainty.
Problem
Existing graph mixup methods interpolate node-centric features, labels, and edges, which can distort graph topology, and they use a randomly sampled, one-size-fits-all mixing ratio for all pairs, ignoring pair-specific needs in semi-supervised node classification.
Approach
Treats each labeled node's r-hop ego subgraph as an analogue of an image, mixes pairs of subgraphs via a virtual node connected to both neighborhoods, and sets lambda per pair using embedding similarity (initialization) and the difference in predictive entropy (uncertainty adjustment).
Key results
- With GCN, AGMixup reaches 83.61% on Cora versus 80.51% for vanilla GCN and at most 81.21% for the other mixup baselines, with an average gain of 2.71% across six datasets.
- AGMixup improves GraphSAGE on Cora to 82.67% and on Pubmed to 79.84%, with average gain 3.00% over backbone.
- On ogbn-products, AGMixup raises GCN test accuracy from 70.48% to 72.22%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply AGMixup instead of vanilla GCN | GCN · Cora | accuracy: 80.51 -> 83.61 | Helped |
| Apply AGMixup instead of the strongest listed mixup baseline, NodeMixup | GCN · Citeseer | accuracy: 69.45 -> 70.63 | Helped |
| Apply AGMixup instead of the strongest listed mixup baseline, NodeMixup | GCN · Pubmed | accuracy: 78.98 -> 81.90 | Helped |
| Apply AGMixup instead of the strongest listed mixup baseline, NodeMixup | GCN · Coauthor CS | accuracy: 92.03 -> 92.35 | Helped |
| Apply AGMixup instead of the strongest listed mixup baseline, NodeMixup | GCN · Coauthor Physics | accuracy: 93.86 -> 94.45 | Helped |
| Apply AGMixup instead of the strongest listed mixup baseline, NodeMixup | GCN · ogbn-arxiv | accuracy: 66.82 -> 67.96 | Helped |
| Replace subgraph-centric mixup with node-centric mixup (r=0) (ablation) | AGMixup · Pubmed | accuracy | Hurt |
| Replace adaptive λ with randomly sampled Beta-distribution λ (ablation) | AGMixup · Cora, Citeseer, Pubmed | accuracy | Hurt |
| Replace adaptive λ with fixed λ=0.5 (ablation) | AGMixup · Cora, Citeseer, Pubmed | accuracy | Hurt |
| Add contextual similarity-aware λ initialization (ablation) | AGMixup · Cora, Citeseer, Pubmed | accuracy | Helped |
| Add uncertainty-aware λ adjustment (ablation) | AGMixup · Cora, Citeseer, Pubmed | accuracy | Helped |
| Increase subgraph radius r up to 3 (ablation) | AGMixup with GCN · Cora | accuracy: 83.90% | Helped |
Surprising findings
- iGraphMix and GraphMix give negligible or negative gains over the backbone on several GNN/dataset pairs (e.g., GAT with GraphMix on CS drops to 88.43 from 90.06).
Limitations
- Hyperparameters (r, gamma, beta, mu) are tuned by grid search and recommended ranges are dataset-dependent.
- Mixup is evaluated mainly on GCN with limited ablation beyond the four backbones.
Experimental setup
Datasets: Cora Citeseer Pubmed Coauthor CS Coauthor Physics ogbn-arxiv ogbn-products
Benchmarks: Cora Citeseer Pubmed Coauthor CS Coauthor Physics ogbn-arxiv ogbn-products
Metrics: node classification accuracy generalization gap
Baselines: GraphMix NodeMixup iGraphMix DropEdge PairNorm R-GMixup F-GMixup
Builds on: mixup GraphMix iGraphMix NodeMixup
Closest papers by meaning
- iGraphMix: Input Graph Mixup Method for Node Classification (ICLR 2024 · similarity 0.91)
- G-Mixup: Graph Data Augmentation for Graph Classification (ICML 2022 · similarity 0.91)
- IntraMix: Intra-Class Mixup Generation for Accurate Labels and Neighbors (NeurIPS 2024 · similarity 0.90)
- Graph Mixup on Approximate Gromov–Wasserstein Geodesics (ICML 2024 · similarity 0.88)
- NodeMixup: Tackling Under-Reaching for Graph Neural Networks (AAAI 2024 · similarity 0.88)
- Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization (AAAI 2024 · 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.