Research map / Graph learning / GNN architectures
Curriculum Learning for Graph Neural Networks: Which Edges Should We Learn First
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 |
|---|---|---|---|
| apply RCL edge curriculum to GCN | GCN · real-world node classification (nine datasets) | accuracy: +1.86% on average over baseline models | Helped |
| apply RCL edge curriculum to GIN | GIN · real-world node classification (nine datasets) | accuracy: +2.83% on average over baseline models | Helped |
| apply RCL edge curriculum to GraphSAGE | GraphSAGE · real-world node classification (nine datasets) | accuracy: +1.62% on average over baseline models | Helped |
| use learned pacing instead of linear edge pacing (ablation) | RCL · ablation datasets | accuracy: +1.59% on average across datasets versus Curriculum-linear and Curriculum-root | Helped |
| use learned pacing instead of root edge pacing (ablation) | RCL · ablation datasets | accuracy: +1.59% on average across datasets versus Curriculum-linear and Curriculum-root | Helped |
| use difficulty-based edge selection instead of random edge selection (ablation) | RCL · ablation datasets | accuracy: 3.86% on average versus Random-linear and Random-root | Helped |
| add edge appearance smoothing (ablation) | RCL · Synthetic1 (homo=0.3) | accuracy: 70.70$\pm$0.43 -> 73.98$\pm$0.55 | Helped |
| add node-confidence weighting (ablation) | RCL · Synthetic1 (homo=0.3) | accuracy: 72.38$\pm$0.41 -> 73.98$\pm$0.55 | Helped |
| initialize the training structure using a pre-trained model (ablation) | RCL · Synthetic2 (homo=0.6) | accuracy: 93.89$\pm$0.14 -> 97.42$\pm$0.17 | Helped |
| add the edge curriculum (ablation) | vanilla GNN · ablation datasets | accuracy: +1.92% on average | Helped |
| use RCL with random structural noise | GCN and GIN · Cora and Citeseer under 100% added-edge attack | accuracy: 4.43%, 2.83% on Cora; 6.13%, 3.47% on Citeseer, with GCN and GIN respectively | Helped |
Closest papers by meaning
- Learning to Pre-train Graph Neural Networks (AAAI 2021 · similarity 0.90)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.89)
- How Graph Neural Networks Learn: Lessons from Training Dynamics (ICML 2024 · similarity 0.88)
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.88)
- Curriculum-Enhanced Residual Soft An-Isotropic Normalization for Over-Smoothness in Deep GNNs (AAAI 2024 · similarity 0.88)
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.88)
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.