Research map / Graph learning / Graph self-supervised learning
Perfect Alignment May be Poisonous to Graph Contrastive Learning
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.
Argues that perfect alignment hurts generalization in graph contrastive learning and proposes information-based and spectrum-based augmentation methods that are added to GRACE, GCA and AD-GCL to improve node classification.
Problem
Understanding how augmentation magnitude affects graph contrastive learning and downstream generalization, and why perfect alignment does not guarantee the best downstream node classification accuracy.
Approach
Downstream gains from stronger augmentation come mainly from inter-class separation rather than intra-class gathering; perfect alignment raises generalization gap. Two plug-in augmentations are derived from theory: an information-based one (keep high-gradient-importance edges/features, drop low-importance ones more aggressively) and a spectrum-based one (adaptively shrink adjacency eigenvalue magn
Key results
- GRACE+I achieves 84.28 mean accuracy vs 83.25 for GRACE across six datasets, and GRACE+S reaches 84.40 with p=0.003.
- GCA+I reaches 84.46 mean accuracy vs 83.77 for GCA (p=0.089), with weaker gains for GCA because it already drops unimportant edges/features.
- AD-GCL+I improves mean accuracy from 82.62 to 83.76 (p=0.003), with the largest gains on AD-GCL which aggressively drops information.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add information augmentation | GRACE · Mean across six datasets | accuracy: 83.25 -> 84.28 | Helped |
| add spectrum augmentation | GRACE · Mean across six datasets | accuracy: 83.25 -> 84.40 | Helped |
| add information augmentation | GCA · Mean across six datasets | accuracy: 83.77 -> 84.46 | Helped |
| add spectrum augmentation | GCA · Mean across six datasets | accuracy: 83.77 -> 84.29 | Helped |
| add information augmentation | AD-GCL · Mean across six datasets | accuracy: 82.62 -> 83.76 | Helped |
| add spectrum augmentation | AD-GCL · Mean across six datasets | accuracy: 82.62 -> 83.74 | Helped |
| increase augmentation strength beyond an appropriate level | graph contrastive learning · Downstream tasks | accuracy | Hurt |
Surprising findings
- Perfect alignment, which is usually assumed beneficial, can hurt generalization; better downstream accuracy is correlated with stronger augmentation and a larger positive center distance rather than intra-class gathering.
- Augmentation overlap between intra-class samples is rare, so the success of contrastive learning cannot be attributed solely to intra-class gathering.
Limitations
- The information-based method randomly selects edges to delete and sets the deletion count to twice the number kept, which the authors note is a simple, theory-verification design rather than tuned.
- Gains for GCA are small and not statistically significant (p=0.089 for GCA+I, 0.147 for GCA+S).
- Evaluation is limited to six node-level datasets and three GCL backbones.
Experimental setup
Datasets: Cora CiteSeer PubMed DBLP Amazon-Photo Amazon-Computers
Benchmarks: Cora CiteSeer PubMed DBLP Amazon-Photo Amazon-Computers
Metrics: node classification accuracy p-value
Baselines: GRACE GCA AD-GCL Supervised GCN Supervised GAT GRACE+SpCo GCS
Builds on: GRACE GCA AD-GCL InfoMin alignment and uniformity
Closest papers by meaning
- Community-Invariant Graph Contrastive Learning (ICML 2024 · similarity 0.91)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.91)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.91)
- Unified Graph Augmentations for Generalized Contrastive Learning on Graphs (NeurIPS 2024 · similarity 0.91)
- A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
- Graph Contrastive Invariant Learning from the Causal Perspective (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.