Research map / Graph learning / Graph self-supervised learning
Contrastive Self-supervised Learning for Graph Classification
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 |
|---|---|---|---|
| pretrain the graph encoder with graph-level contrastive learning, then fine-tune it | HGP-SL graph classification · PROTEINS, D&D, NCI1, NCI109, Mutagenicity | accuracy (%): 85.71$\pm$0.69; 82.02$\pm$1.42; 78.62$\pm$0.63; 80.72$\pm$1.06; 82.00$\pm$0.63 | Helped |
| jointly optimize classification and the graph contrastive task as a regularizer | HGP-SL graph classification · PROTEINS, D&D, NCI1, NCI109, Mutagenicity | accuracy (%): 84.11$\pm$0.87; 82.18$\pm$1.34; 80.07$\pm$0.60; 81.16$\pm$1.42; 82.07$\pm$0.65 | Helped |
| fine-tune the pretrained encoder instead of freezing it (ablation) | CSSL-Pretrain · graph classification datasets | accuracy (%) | Helped |
| use three consecutive alterations instead of one (ablation) | CSSL-Pretrain and CSSL-Reg · graph classification datasets | accuracy (%) | Mixed |
| define CSSL on the target dataset rather than training graphs from all five datasets (ablation) | CSSL-Pretrain and CSSL-Reg · graph classification datasets | accuracy (%) | Helped |
| use random alteration types instead of deterministically selecting one type (ablation) | CSSL-Reg · PROTEINS, D&D, NCI1, NCI109 | accuracy (%): 84.1 vs 78.4; 82.2 vs 80.3; 80.1 vs 77.5; 81.2 vs 79.1 | Helped |
| increase the CSSL-Reg regularization weight from zero, then avoid making it too large (ablation) | CSSL-Reg · graph classification datasets | accuracy | Mixed |
| replace graph-level contrastive learning with local self-supervised representations | graph classification · five graph classification datasets | accuracy (%) | Hurt |
Closest papers by meaning
- Simple and Asymmetric Graph Contrastive Learning without Augmentations (NeurIPS 2023 · similarity 0.90)
- Decoupled Self-supervised Learning for Graphs (NeurIPS 2022 · similarity 0.89)
- Improving Graph Contrastive Learning via Adaptive Positive Sampling (CVPR 2024 · similarity 0.88)
- A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning (AAAI 2024 · similarity 0.87)
- Equivalence is All: A Unified View for Self-supervised Graph Learning (ICML 2025 · similarity 0.87)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.87)
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.