Research map / Graph learning / Graph self-supervised learning
Mitigating Local Cohesion and Global Sparseness in Graph Contrastive Learning with Fuzzy Boundaries
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 a fuzzy boundary (fuzzy set construction and contraction) on a pre-trained GCN graph auto-encoder to fix local cohesion and global sparseness in graph contrastive learning.
Problem
Graph contrastive learning produces embeddings where class clusters fragment into many small groups and isolated samples, hurting intra-cluster compactness and inter-cluster separation.
Approach
Extends original cluster boundaries with learnable Gaussian fuzzy membership functions (max-pooling union) to build fuzzy boundary elements that bridge local groups, then contracts them toward fused prototypes via a prototype contrastive loss, fine-tuning a pre-trained GAE.
Key results
- Node classification: achieves 85.9 on Cora, 74.1 on Citeseer, 87.0 on Pubmed, 69.5 on Ogbn-arxiv, and best results on most datasets versus baselines such as PiGCL and SGRL.
- Node clustering on Citeseer: Acc 70.25 and NMI 44.99; on Amazon Photo Acc 76.71 and NMI 67.72.
- DBSCAN quantitative analysis (Eps=0.7, Citeseer): Mean Groups 1.7 and Total Isolated Nodes 78, versus 3.3 and 208 for SGRL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the fuzzy-boundary model instead of GREET | Node classification · Cora | accuracy: 85.7 ± 0.5 -> 85.9 ± 0.6 | Helped |
| Use the fuzzy-boundary model instead of SUGRL | Node classification · Citeseer | accuracy: 73.5 ± 0.6 -> 74.1 ± 0.6 | Helped |
| Use the fuzzy-boundary model instead of GREET | Node classification · Pubmed | accuracy: 86.9 ± 0.3 -> 87.0 ± 0.3 | Helped |
| Use the fuzzy-boundary model instead of COAST | Node classification · DBLP | accuracy: 84.5 ± 0.1 -> 84.8 ± 0.2 | Helped |
| Use the fuzzy-boundary model instead of SUGRL | Node classification · Ogbn-arxiv | accuracy: 68.7 ± 0.0 -> 69.5 ± 0.1 | Helped |
| Use the fuzzy-boundary model instead of SGRL | Node classification · Amazon Photo | accuracy: 93.8 ± 0.0 -> 93.7 ± 0.2 | Hurt |
| Use the fuzzy-boundary model instead of iGCL | Node classification · Amazon Computers | accuracy: 90.1 ± 0.4 -> 90.8 ± 0.2 | Helped |
| Use the fuzzy-boundary model instead of GREET | Node clustering · Cora | Acc: 71.83 ± 3.40 -> 72.85 ± 0.02 | Helped |
| Use the fuzzy-boundary model instead of GREET | Node clustering · Citeseer | Acc: 68.78 ± 0.18 -> 70.25 ± 0.09 | Helped |
| Use the fuzzy-boundary model instead of HomoGCL | Node clustering · Amazon Photo | Acc: 74.30 ± 1.69 -> 76.71 ± 0.06 | Helped |
| Use the fuzzy-boundary model instead of iGCL | Node clustering · Amazon Computers | Acc: 53.49 ± 2.47 -> 55.04 ± 0.78 | Helped |
| Use the fuzzy-boundary model instead of HomoGCL | Node clustering · Cora | MaF1: 66.10 ± 5.87 -> 71.50 ± 0.01 | Helped |
Surprising findings
- GREET reports OOM on Ogbn-arxiv, PiGCL OOM on Ogbn-arxiv, showing scalability issues of some contrastive baselines.
Limitations
- Results on fuzzy boundary depend on K-means clustering of pre-trained representations, which is run every T epochs.
- Theoretical claims (generalization improvement) are provided via appendix proofs only and experiments are limited to seven graph datasets.
Experimental setup
Datasets: Cora Citeseer Pubmed DBLP Ogbn-arxiv Amazon Photo Amazon Computers
Benchmarks: Cora Citeseer Pubmed DBLP Ogbn-arxiv Amazon Photo Amazon Computers
Metrics: accuracy macro-F1 NMI ARI Mean Groups Total Isolated Nodes
Baselines: DeepWalk DGI GMI GRACE GCA COAST SUGRL AFGRL iGCL NCLA GREET HomoGCL
Builds on: graph contrastive learning fuzzy set theory graph auto-encoder DeepWalk
Closest papers by meaning
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.87)
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.87)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.87)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.87)
- THESAURUS: Contrastive Graph Clustering by Swapping Fused Gromov-Wasserstein Couplings (AAAI 2025 · similarity 0.86)
- StructComp: Substituting propagation with Structural Compression in Training Graph Contrastive Learning (ICLR 2024 · similarity 0.86)
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.