Research map / Graph learning / Graph self-supervised learning
THESAURUS: Contrastive Graph Clustering by Swapping Fused Gromov-Wasserstein Couplings
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 THESAURUS, a contrastive graph clustering method that uses semantic prototypes, cross-view swapped assignment prediction, and fused Gromov-Wasserstein OT with momentum-updated prototype graph and marginal to improve cluster separability.
Problem
Deep graph node clustering methods that rely on K-means suffer from low cluster separability, causing the Uniform Effect (majority clusters absorbing minority-cluster samples) and Cluster Assimilation (minority clusters merged into majority ones).
Approach
Introduces learnable semantic prototypes that give each node a context-aware representation (similarities to prototypes), trains with a swapped cross-view assignment task solved by fused Gromov-Wasserstein OT between the data graph and a momentum-updated prototype graph, and runs K-means on the context-aware representations.
Key results
- On Pubmed, THESAURUS reaches 79.64 ACC and 48.25 ARI versus 68.73/29.84 for the best-listed baseline without Dink-Net... and 67.51 ACC for Dink-Net.
- On ACM, DBLP, UAT and Wiki (5 runs), THESAURUS reports ACC 92.52, 82.25, 62.25, 57.84, outperforming Dink-Net (90.54, 74.71, 58.07, 53.26).
- Macro-F1 on Cora is up to 5.5 points higher than Dink-Net and 11.65 points higher on Pubmed; on Cora THESAURUS ACC 80.72 vs Dink-Net 78.21.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the complete prototype graph instead of using an isolated graph (ablation) | THESAURUS · Citeseer | F1: 65.02 -> 66.42 | Helped |
| add the complete prototype graph instead of using an isolated graph (ablation) | THESAURUS · Pubmed | F1: 75.27 -> 79.00 | Helped |
| retain data graph structure in FGW-OT instead of replacing adjacency with an identity matrix (ablation) | THESAURUS · Citeseer | F1: 65.75 -> 66.42 | Helped |
| retain data graph structure in FGW-OT instead of replacing adjacency with an identity matrix (ablation) | THESAURUS · Pubmed | F1: 77.81 -> 79.00 | Helped |
| update the prototype graph and marginal with momentum rather than fixing them (ablation) | THESAURUS · Citeseer | F1: 65.68 -> 66.42 | Helped |
| update the prototype graph and marginal with momentum rather than fixing them (ablation) | THESAURUS · Pubmed | F1: 75.39 -> 79.00 | Helped |
| use THESAURUS instead of Dink-Net | graph clustering · Cora | F1: 72.85 -> 78.44 | Helped |
| use THESAURUS instead of Dink-Net | graph clustering · Citeseer | F1: 65.79 -> 66.42 | Helped |
| use THESAURUS instead of Dink-Net | graph clustering · Pubmed | F1: 67.35 -> 79.00 | Helped |
| use THESAURUS instead of Dink-Net | graph clustering · A-Photo | F1: 73.09 -> 76.49 | Helped |
| use THESAURUS instead of Dink-Net | graph clustering · CoraFull | F1: 31.95 -> 37.01 | Helped |
Surprising findings
- Pubmed ACC for SCGC drops to 45.12 and ARI to 7.04 in the reported table, showing unstable baselines on that dataset.
Limitations
- Some baselines (HSAN, S3GC-style methods) run out of memory on larger datasets (Pubmed, CoraFull), so comparisons are incomplete there.
- Main-text results are single runs; multi-run statistics only given for four appendix datasets.
Experimental setup
Datasets: Cora Citeseer Pubmed Amazon-Photo CoraFull ACM DBLP UAT Wiki
Benchmarks: Cora Citeseer Pubmed Amazon-Photo CoraFull ACM DBLP UAT Wiki
Metrics: ACC NMI ARI F1
Baselines: K-means DEC SDCN GRACE DAEGC DFCN DCRN HSAN S3GC SCGC Dink-Net
Builds on: Dink-Net SwAv Fused Gromov-Wasserstein OT Sinkhorn GRACE
Closest papers by meaning
- Cluster-Guided Contrastive Graph Clustering Network (AAAI 2023 · similarity 0.90)
- GLCC: A General Framework for Graph-Level Clustering (AAAI 2023 · similarity 0.90)
- Discriminative Attribute Graph Clustering Through Topology-Guided Contrastive Learning (ICML 2026 · similarity 0.90)
- Hybrid-Collaborative Augmentation and Contrastive Sample Adaptive-Differential Awareness for Robust Attributed Graph Clustering (NeurIPS 2025 · similarity 0.88)
- Multi-View Graph Clustering via Node-Guided Contrastive Encoding (ICML 2025 · similarity 0.88)
- Deep Contrastive Graph Learning with Clustering-Oriented Guidance (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.