Research map / Graph learning / Graph self-supervised learning
Architecture Matters: Uncovering Implicit Mechanisms in Graph Contrastive Learning
Xiaojun Guo, Yifei Wang, Zeming Wei, Yisen Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove positive pairs (uniformity loss only) (ablation) | SimCLR · CIFAR-10 linear probing | accuracy (%): 83.51 -> 27.51 | Hurt |
| remove positive pairs (uniformity loss only) (ablation) | GRACE · Node classification avg, 5 datasets | accuracy (%): 84.87 -> 83.43 | No clear effect |
| remove negative pairs (alignment loss only) (ablation) | GRACE · Node classification avg, 5 datasets | accuracy (%): 84.87 -> 30.37 | Hurt |
| remove negative pairs (alignment loss only) (ablation) | GraphCL · Graph classification avg, 6 datasets | accuracy (%): 71.29 -> 70.77 | No clear effect |
| remove negative pairs (alignment loss only) (ablation) | AutoGCL · Graph classification avg, 6 datasets | accuracy (%): 69.87 -> 70.60 | No clear effect |
| replace GCN encoder with MLP under no-positive training (ablation) | GRACE · Node classification avg, 5 datasets | accuracy (%): 77.07 -> 62.77 | Hurt |
| add ContraNorm to GCN encoder with alignment loss only (ablation) | GRACE · Node classification avg, 5 datasets | accuracy (%): 30.37 -> 82.96 | Improved |
| replace FM+EP augmentation with random Gaussian noise (ablation) | GRACE · Node classification avg, 5 datasets | accuracy (%): 84.87 -> 82.95 | No clear effect |
| remove all augmentations (ablation) | GRACE · Node classification avg, 5 datasets | accuracy (%): 84.87 -> 81.98 | No clear effect |
| replace default augmentation with Gaussian noise (ablation) | SimCLR · CIFAR-10 linear probing | accuracy (%): 83.51 -> 36.56 | Hurt |
| remove positive pairs on heterophilic graphs (ablation) | GRACE · Chameleon and Squirrel avg | accuracy (%): 40.88 -> 36.73 | Hurt |
| remove positive pairs (uniformity loss only) (ablation) | GRACE · OGB-arxiv | accuracy (%): 65.97 -> 65.49 | No clear effect |
Closest papers by meaning
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.92)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.92)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.92)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.91)
- GNN-Transformer Cooperative Architecture for Trustworthy Graph Contrastive Learning (AAAI 2025 · similarity 0.90)
- A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
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.
