Research map / Graph learning / Graph self-supervised learning
Exploitation of a Latent Mechanism in Graph Contrastive Learning: Representation Scattering
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.
Shows that node-discrimination, group-discrimination and bootstrapping graph contrastive learning implicitly perform representation scattering, and proposes SGRL, which adds a center-away scattering loss and a topology-based constraint with EMA target updates.
Problem
Graph contrastive learning has three mainstream frameworks (InfoNCE-based, DGI-like, BGRL-like) with comparable performance, but the shared mechanism behind their success is not understood, and existing methods implement it inefficiently or with false negatives.
Approach
Defines representation scattering (center-away plus uniformity on a hypersphere), argues each GCL framework implicitly performs it, and builds SGRL with a direct hypersphere center-away loss (RSM) plus a k-hop topology-aligned predictor constraint (TCM) and EMA target encoder, avoiding explicit negative samples.
Key results
- SGRL reaches 79.40 F1 on WikiCS, 90.23 on Computers, 93.95 on Photo, 94.15 on Co-CS and 96.23 on Co-Physics, the best among compared methods on all five datasets.
- Compared with BGRL, GRACE and DGI, SGRL improves average node classification by 1.23%, 2.14% and 3.26% respectively.
- Removing batch normalization from BGRL drops its F1-score across Photo, Co.CS, Computers and Physics (Figure 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add RSM and TCM to the SGRL encoder (ablation) | SGRL · WikiCS | F1-score: 76.86 ± 0.74 -> 79.40 ± 0.13 | Helped |
| Add RSM and TCM to the SGRL encoder (ablation) | SGRL · Computers | F1-score: 89.69 ± 0.37 -> 90.23 ± 0.03 | Helped |
| Add RSM and TCM to the SGRL encoder (ablation) | SGRL · Photo | F1-score: 93.07 ± 0.38 -> 93.95 ± 0.03 | Helped |
| Add RSM and TCM to the SGRL encoder (ablation) | SGRL · Co.CS | F1-score: 92.59 ± 0.14 -> 94.15 ± 0.04 | Helped |
| Add RSM and TCM to the SGRL encoder (ablation) | SGRL · Co-Physics | F1-score: 95.48 ± 0.08 -> 96.23 ± 0.01 | Helped |
| Add TCM to SGRL with RSM (ablation) | SGRL · WikiCS | F1-score: 78.55 ± 0.08 -> 79.40 ± 0.13 | Helped |
| Add TCM to SGRL with RSM (ablation) | SGRL · Computers | F1-score: 89.54 ± 0.10 -> 90.23 ± 0.03 | Helped |
| Add TCM to SGRL with RSM (ablation) | SGRL · Photo | F1-score: 93.58 ± 0.05 -> 93.95 ± 0.03 | Helped |
| Add TCM to SGRL with RSM (ablation) | SGRL · Co.CS | F1-score: 94.08 ± 0.03 -> 94.15 ± 0.04 | Helped |
| Add TCM to SGRL with RSM (ablation) | SGRL · Co-Physics | F1-score: 96.19 ± 0.04 -> 96.23 ± 0.01 | Helped |
| Add EMA to balance RSM and TCM (ablation) | SGRL · WikiCS | F1-score: 79.36 ± 0.08 -> 79.40 ± 0.13 | Helped |
| Add EMA to balance RSM and TCM (ablation) | SGRL · Computers | F1-score: 90.03 ± 0.07 -> 90.23 ± 0.03 | Helped |
Surprising findings
- Removing the topology-based constraint (TCM) slightly lowers performance on Amazon-Computers relative to RSM alone, which the authors attribute to over-scattering.
Limitations
- Evaluated only on homogeneous graph datasets; applicability to heterogeneous graphs is untested.
- Balancing RSM and TCM relies on EMA, which is only one possible balancing method.
- Theory assumes class-balanced data; imbalanced classes shift the scattering center.
Experimental setup
Datasets: WikiCS Amazon-Computers Amazon-Photo Coauthor-CS Coauthor-Physics
Benchmarks: WikiCS Amazon-Computers Amazon-Photo Coauthor-CS Coauthor-Physics
Metrics: F1-score NMI homogeneity
Baselines: GRACE DGI BGRL GCA ProGCL AFGRL iGCL GBT MVGRL Node2vec DeepWalk GCN
Builds on: BGRL DGI GRACE BYOL
Closest papers by meaning
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.91)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.90)
- Hierarchical Topology Isomorphism Expertise Embedded Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.89)
- Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing (AAAI 2025 · similarity 0.89)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.89)
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.