Research map / Graph learning / Graph self-supervised learning
Edge Contrastive Learning: An Augmentation-Free Graph Contrastive Learning Model
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 AFECL, an augmentation-free graph contrastive learning method that builds edge embeddings by concatenating connected node embeddings and contrasts edges sharing a node as positives, evaluated on node classification with scarce labels and link prediction.
Problem
Existing graph contrastive learning relies on handcrafted augmentations that can damage topology and uses node-level objectives that ignore edge structure and the homophily assumption; edge-level contrast is underexplored due to its computational cost.
Approach
Generate edge features by concatenating (linearly projecting) the GAT node embeddings of their two endpoints, and define an edge-edge contrastive loss where edges sharing an endpoint with the anchor (plus the anchor itself) are positives and edges not touching its neighborhood are negatives, with no augmentation and a single view.
Key results
- With 1 label per class, AFECL reaches 63.6% on Cora and 78.9% on Coauthor-CS, the best among compared methods in the paper's table.
- On link prediction, AFECL reports AUC 96.5 on Cora, 96.6 on CiteSeer and 98.2 on Coauthor-CS, the highest among the compared methods.
- On ogbn-arxiv, AFECL reaches 72.5±0.2% versus 72.2±0.1% for CCA-SSG, using a single-view setting with compressed graphs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use edge contrastive learning instead of the w/o ECL loss variant (ablation) | AFECL · Cora, 1 label/class | accuracy: 48.1 -> 63.6 | Helped |
| use edge contrastive learning instead of the w/o ECL loss variant (ablation) | AFECL · Citeseer, 1 label/class | accuracy: 41.7 -> 50.9 | Helped |
| use edge contrastive learning instead of the w/o ECL loss variant (ablation) | AFECL · PubMed, 1 label/class | accuracy: 45.1 -> 59.8 | Helped |
| use edge contrastive learning instead of the w/o ECL loss variant (ablation) | AFECL · Coauthor-CS, 1 label/class | accuracy: 50.7 -> 78.9 | Helped |
| use edge contrastive learning instead of the w/o ECL loss variant (ablation) | AFECL · Amazon-Photo, 1 label/class | accuracy: 54.2 -> 73.9 | Helped |
| use edge contrastive learning instead of the w/o ECL loss variant (ablation) | AFECL · Actor, 1 label/class | accuracy: 20.9 -> 21.9 | Helped |
| use edge contrastive learning instead of the w/o ECL loss variant (ablation) | AFECL · Chameleon, 1 label/class | accuracy: 26.4 -> 26.9 | Helped |
| use edge contrastive learning instead of the w/o ECL loss variant (ablation) | AFECL · node classification across the seven datasets in the ablation table | accuracy | Helped |
| use AFECL rather than the strongest listed baseline | semi-supervised node classification · Cora, Citeseer, PubMed, Coauthor-CS, and Amazon-Photo with 20 labels/class | accuracy | Mixed |
| use AFECL rather than the strongest listed baseline | link prediction · Cora, Citeseer, and Coauthor-CS | area under the ROC curve | Helped |
| use AFECL rather than HLCL | Penn94 · Penn94 | test accuracy: 68.1±3.5 -> 68.2±3.0 | No clear effect |
| use AFECL rather than the strongest listed baseline | ogbn-arxiv · ogbn-arxiv | test accuracy: 72.2±0.1 -> 72.5±0.2 | Helped |
Surprising findings
- Defining edges that share only one endpoint as negatives (w/o ECL variant) drops accuracy substantially, e.g., Cora with 1 label per class falls from 63.6% to 48.1%.
Limitations
- Edge contrast cost scales with the square of the number of sampled edges, requiring edge sampling on large graphs.
- Relies on single-view setting with no augmentation; benefits on heterophilic graphs are small (Actor accuracy around 22-25%).
Experimental setup
Datasets: Cora CiteSeer PubMed Coauthor-CS Amazon-Photo Actor Chameleon Penn94 ogbn-arxiv
Benchmarks: Cora CiteSeer PubMed Coauthor-CS Amazon-Photo Actor Chameleon Penn94 ogbn-arxiv
Metrics: node classification accuracy link prediction AUC
Baselines: GCN GAT CCA-SSG DGI BGRL MVGRL GRACE GCA ARIEL SPGCL GraphACL PiGCL
Builds on: GAT GRACE GraphACL InfoNCE
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 0.90)
- Unified Graph Augmentations for Generalized Contrastive Learning on Graphs (NeurIPS 2024 · similarity 0.90)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.89)
- Why Does Dropping Edges Usually Outperform Adding Edges in Graph Contrastive Learning? (AAAI 2025 · similarity 0.89)
- Improving Graph Contrastive Learning via Adaptive Positive Sampling (CVPR 2024 · similarity 0.89)
- Graph Contrastive Learning with Joint Spectral Augmentation of Attribute and Topology (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.