Research map / Graph learning / Graph self-supervised learning
Unified Graph Augmentations for Generalized Contrastive Learning on Graphs
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 UGA, a unified graph augmentation module that simulates node, edge, attribute and subgraph augmentations via shared augmentation-centric vectors propagated by attention, and GOUDA, a contrastive framework combining InfoNCE or Barlow Twins with an HSIC-based independence loss.
Problem
Existing graph contrastive learning augmentations are task-specific, computationally expensive (e.g., spectral, adversarial, set-sampling methods), and limited in the range of augmentations they can express.
Approach
Reinterpret all explicit graph augmentations as local attribute modifications in node neighborhoods from a message-passing view, so a small set of learnable augmentation-centric vectors added to node features can emulate any of them; then train two views with a contrastive loss for consistency plus an HSIC independence loss between the two augmentation matrices for diversity.
Key results
- On CiteSeer node classification, GOUDA-IF reaches 74.55% vs. GRACE's 71.41% (about 3.14 points gain), and GOUDA-BT beats GBT by 1.90 points.
- On IMDB-B graph classification, GOUDA-BT reaches 76.80% and GOUDA-IF 75.22%, surpassing the second-best MVGRL (74.20%).
- GOUDA-IF achieves the best NMI/ARI on Cora (57.92/52.41) and PubMed clustering, outperforming CCA-SSG and GBT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GOUDA-IF instead of GRACE | Node classification · CiteSeer | accuracy: +3.14% | Helped |
| Use GOUDA-BT instead of GBT | Node classification · CiteSeer | accuracy: +1.90% | Helped |
| Use GOUDA-IF instead of MVGRL | Graph classification · IMDB-B | accuracy: 1.02% | Helped |
| Use GOUDA-BT instead of MVGRL | Graph classification · IMDB-B | accuracy: 2.6% | Helped |
| Remove structural features from UGA (ablation) | GOUDA · Node classification | accuracy | Hurt |
| Add structural features to UGA (ablation) | GOUDA · Node classification | accuracy | Helped |
| Remove the independence loss (ablation) | GOUDA · Node classification | accuracy | Hurt |
| Add the independence loss (ablation) | GOUDA · Node classification | accuracy | Helped |
| Use InfoNCE rather than Barlow Twins (ablation) | GOUDA without independence loss · Node classification | accuracy | Hurt |
| Vary the number of AC vectors across 5, 10, and 20 (ablation) | GOUDA · Node classification | accuracy: within a 2% margin | No clear effect |
Surprising findings
- Without the independence loss, GOUDA-IF performs worse than GOUDA-BT, which the authors attribute to InfoNCE driving excessive consistency.
- Attribute attacks cause clearly larger performance degradation than topology attacks, even for GOUDA.
Limitations
- Robustness against attribute attacks is weaker than against topology attacks.
- Theoretical equivalence relies on linear one-layer encoders without activations and small-perturbation approximations for the diversity bound.
Experimental setup
Datasets: Cora CiteSeer PubMed Wiki-CS Amazon Photo Amazon Computers Coauthor Physics IMDB-B IMDB-M COLLAB
Benchmarks: Cora CiteSeer PubMed Wiki-CS Photo Computers Physics IMDB-B IMDB-M COLLAB
Metrics: accuracy NMI ARI running time GPU memory
Baselines: GCN GAT DGI GMI MVGRL GRACE GCA BGRL GBT CCA-SSG SPAN DSSL
Builds on: GRACE BGRL GBT SPAN AD-GCL JOAO
Closest papers by meaning
- Graph Contrastive Learning with Joint Spectral Augmentation of Attribute and Topology (AAAI 2025 · similarity 0.91)
- Community-Invariant Graph Contrastive Learning (ICML 2024 · similarity 0.91)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.91)
- Edge Contrastive Learning: An Augmentation-Free Graph Contrastive Learning Model (AAAI 2025 · similarity 0.90)
- GNN-Transformer Cooperative Architecture for Trustworthy Graph Contrastive Learning (AAAI 2025 · similarity 0.89)
- SOLA-GCL: Subgraph-Oriented Learnable Augmentation Method for Graph Contrastive Learning (AAAI 2025 · similarity 0.89)
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.