Research map / Graph learning / Graph self-supervised learning
An Empirical Study of Graph Contrastive Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use topology augmentations that create sparse views rather than adding edges | GCL · Node and graph classification | accuracy | Helped |
| Use RWS rather than other tested topology augmentations | GCL · Node classification | accuracy | Helped |
| Use node dropping rather than other tested topology augmentations | GCL · Graph classification | accuracy | Helped |
| Combine feature augmentation with topology augmentation | GCL · Node and graph classification | accuracy | Helped |
| Use a shared feature mask (FM) instead of feature dropout (FD) | GCL · Node and graph classification | accuracy | No clear effect |
| Pair deterministic diffusion augmentations with stochastic augmentations | GCL · Node and graph classification | accuracy | Helped |
| Use same-scale contrasting, with local-local for node tasks and global-global for graph tasks | GCL · Node and graph classification | accuracy | Helped |
| Use InfoNCE rather than other negative-sample-based objectives | GCL · Node and graph classification | accuracy | Helped |
| Use negative-sample-free BL or BT objectives instead of negative-sample-based objectives | GCL · Node and graph classification | accuracy | Mixed |
| Mine hard negatives based on embedding similarity | InfoNCE GCL · Node classification | accuracy | No clear effect |
| Add batch normalization to the encoder (ablation) | Bootstrapping Latent · Node classification | accuracy | Helped |
| Remove batch normalization from the encoder while retaining it in projector and predictor (ablation) | Bootstrapping Latent · Wiki, CS, Physics, and Computer | accuracy: 80.61{ ±0.04} -> 79.42{ ±0.09}; 93.29{ ±0.07} -> 93.53{ ±0.03}; 95.31{ ±0.01} -> 95.23{ ±0.05}; 89.81{ ±0.07} -> 87.45{ | Hurt |
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 0.94)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.93)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.93)
- Hierarchical Topology Isomorphism Expertise Embedded Graph Contrastive Learning (AAAI 2024 · similarity 0.92)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.92)
- A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning (AAAI 2024 · similarity 0.91)
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.