Research map / Graph learning / Graph self-supervised learning
ML$^2$-GCL: Manifold Learning Inspired Lightweight Graph Contrastive Learning
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 ML2-GCL, a single-view graph contrastive learning method that uses a closed-form, locally linear manifold-learning reconstruction to set positive-pair weights and uniform negative weights, encoded by a GCN.
Problem
Graph contrastive learning relies on augmentations that can disturb semantics and on multi-view or distance-based node sampling that is costly and rigid, while GCNs struggle to capture long-range dependence.
Approach
Replace augmented views and pairwise distance-based positive sampling with LLE-style closed-form reconstruction weights over k-hop neighbors as soft positive-pair weights, with uniform weights on negatives, trained with a single GCN view.
Key results
- Achieves 83.0% node classification accuracy on Amazon-Computers and 73.3% on Wiki-CS, reported as 3.2% and 3.0% relative improvements over prior state of the art.
- Reaches 96.9% AUC on Cora for link prediction, a 3.6% relative improvement over the previous best.
- Ablation shows removing positive-pair weights hurts more than removing negative-pair weights.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ML2-GCL instead of the strongest listed baseline, NCLA | node classification · Amazon-Computers | accuracy: 79.8 ± 1.5 -> 83.0 ± 2.1 | Helped |
| Add positive-pair weights (ablation) | ML2-GCL · Cora node classification | accuracy: 81.9 ± 1.3 -> 83.0 ± 1.3 | Helped |
| Add negative-pair weights (ablation) | ML2-GCL · Cora node classification | accuracy: 82.5 ± 1.2 -> 83.0 ± 1.3 | Helped |
| Add positive-pair weights and negative-pair weights (ablation) | ML2-GCL · Cora node classification | accuracy: 81.2 ± 1.3 -> 83.0 ± 1.3 | Helped |
Surprising findings
- Positive-pair weights matter more than negative-pair weights in the ablation study.
Limitations
- Positive pairs restricted to k-hop neighborhoods, so reliance on graph homophily is implied
- Quadratic-programming weight computation and N^2 complexity terms may limit scaling to very large graphs
Experimental setup
Datasets: Cora Citeseer Pubmed Amazon-Photo Amazon-Computers Wiki-CS
Benchmarks: Cora Citeseer Pubmed Amazon-Photo Amazon-Computers Wiki-CS
Metrics: accuracy AUC AP
Baselines: GCN GAT DGI GMI MVGRL GRACE GCA SUGRL AFGRL NCLA GTCA CGPN
Builds on: GRACE LLE InfoNCE GCN BGRL
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)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.90)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
- TopoGCL: Topological Graph Contrastive Learning (AAAI 2024 · similarity 0.89)
- Hierarchical Topology Isomorphism Expertise Embedded Graph Contrastive Learning (AAAI 2024 · similarity 0.89)
- Improving Graph Contrastive Learning via Adaptive Positive Sampling (CVPR 2024 · 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.