Research map / Graph learning / Graph self-supervised learning
Community-Invariant 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.
Proposes CI-GCL, a learnable graph contrastive learning framework that maximizes spectral changes under a community-invariance constraint, applied uniformly to topology and feature augmentation.
Problem
Random or knowledge-based graph augmentations in graph contrastive learning corrupt high-level community structure and lack unified constraints across topology and node features.
Approach
Maximize graph spectral changes during learnable topology (edge add/drop, node drop) and feature (bipartite feature-node graph) augmentation, which empirically and theoretically preserves community structure while perturbing the graph.
Key results
- Unsupervised TU graph classification average accuracy 77.74% vs. 77.25% for SEGA (second best) across 8 TU datasets.
- Unsupervised OGB regression average RMSE 1.606 vs. 2.184 for GCL-SPAN.
- Transfer learning average ROC-AUC 75.1% vs. 73.6% for SEGA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the community-invariant constraint to topology and feature augmentation | CI-GCL · TU unsupervised graph classification: average | accuracy: 77.25 -> 77.74 | Helped |
| Add the community-invariant constraint to topology and feature augmentation | CI-GCL · Unsupervised graph regression: average | RMSE: 2.184 -> 1.606 | Helped |
| Add the community-invariant constraint to topology and feature augmentation | CI-GCL · 10% labeled semi-supervised graph classification: average | accuracy: 73.1 -> 74.0 | Helped |
| Add the community-invariant constraint to topology and feature augmentation | CI-GCL · Transfer learning: average | ROC-AUC: 73.6 -> 75.1 | Helped |
| Add topology augmentation on one branch (ablation) | CI-GCL · Unsupervised graph classification: average across ablation datasets | accuracy: 79.3 -> 81.5 | Helped |
| Add feature augmentation on one branch (ablation) | CI-GCL · Unsupervised graph classification: average across ablation datasets | accuracy: 79.6 -> 81.5 | Helped |
| Add the community-invariant constraint to topology augmentation (ablation) | CI-GCL · Unsupervised graph classification: average across ablation datasets | accuracy: 80.5 -> 81.5 | Helped |
| Add the community-invariant constraint to feature augmentation (ablation) | CI-GCL · Unsupervised graph classification: average across ablation datasets | accuracy: 80.5 -> 81.5 | Helped |
| Add community-invariant constraints to both topology and feature augmentation (ablation) | CI-GCL · Unsupervised graph classification: average across ablation datasets | accuracy: 79.5 -> 81.5 | Helped |
Surprising findings
- Adding CI constraint to AD-GCL and AutoGCL as plug-in yielded gains on several datasets, though CI+AutoGCL underperforms on PROTEINS and CI+AD-GCL underperforms CI-GCL overall.
- Reported community changes are lowest for CI-GCL, but GraphCL and AutoGCL with larger community disruption underperform.
Limitations
- Eigendecomposition scaling cost O(Mn^3) motivates approximations (Lanczos, truncated SVD) whose effect on quality is not deeply analyzed.
- Community change is measured via spectral clustering, a proxy that depends on clustering choice.
- Gains over baselines are often small relative to standard deviations, e.g., MUTAG where CI-GCL is not best.
Experimental setup
Datasets: NCI1 PROTEINS DD MUTAG COLLAB RDT-B RDT-M5K IMDB-B GITHUB ogbg-molesol ogbg-molipo ogbg-molfreesolv
Benchmarks: TU datasets OGB molecular regression MoleculeNet transfer PPI
Metrics: accuracy RMSE ROC-AUC community changes
Baselines: InfoGraph GraphCL MVGRL JOAO SEGA GCS GCL-SPAN AD-GCL AutoGCL
Builds on: GraphCL AD-GCL GCL-SPAN GIN Gumbel-Softmax
Closest papers by meaning
- Revisiting Graph Contrastive Learning from the Perspective of Graph Spectrum (NeurIPS 2022 · similarity 0.93)
- Spectral Augmentation for Self-Supervised Learning on Graphs (ICLR 2023 · similarity 0.93)
- Graph Contrastive Learning with Joint Spectral Augmentation of Attribute and Topology (AAAI 2025 · similarity 0.92)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.91)
- Unified Graph Augmentations for Generalized Contrastive Learning on Graphs (NeurIPS 2024 · similarity 0.91)
- Hierarchical Topology Isomorphism Expertise Embedded 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.