Research map / Graph learning / GNN architectures
From Local Structures to Size Generalization in Graph Neural Networks
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 |
|---|---|---|---|
| Pattern-tree SSL pretext task (pretraining) added to a 3-layer GNN | real social and biological graph datasets (domain adaptation) · 7 datasets average | test accuracy: 63.5% -> 67.5% | Helped |
| Pattern-tree SSL task used as multi-task learning instead of pretraining | real datasets · 7 datasets average | test accuracy: 63.5% -> 63.6% | Helped |
| Pattern-tree SSL pretraining | datasets with high local-pattern discrepancy (IMDB, Deezer, Proteins, Twitch) · domain-shift social/biological datasets | test accuracy: about 5.4% average gain over vanilla | Helped |
| Pattern-tree SSL pretraining | Deezer · Deezer | test accuracy: 41.1 -> 44.0 is improvement in the table; vanilla 41.1 vs Pattern PT 44.0 | Hurt |
| Scaling train connectivity radius by sqrt(2) to keep expected node degree matched (vs unscaled) (ablation) | first-order GNN on max-clique task · geometric graphs, size generalization | test loss (lower is better): 402 -> 96 (1-layer) | Helped |
| Matching average degree between train and test (p=0.15 vs p=0.3) in G(n,p) (ablation) | first-order GNN student-teacher and edge count tasks · G(n,p), test n=100 | loss: edge count loss 1.8e6 -> 5.1; student-teacher 3500 -> 0.02 (test loss, mean) | Helped |
| Increasing GNN depth from 1 to 3 layers under pattern distribution shift (max clique) (ablation) | first-order GNN · max-clique, rho ratio 1 | test loss: 402 (1-layer) -> 2325 (3-layer) | Hurt |
| Semi-supervised pattern pretraining with 0-5 labeled target examples | real datasets, semi-supervised DA · average over datasets | classification accuracy | Helped |
Closest papers by meaning
- SizeShiftReg: a Regularization Method for Improving Size-Generalization in Graph Neural Networks (NeurIPS 2022 · similarity 0.93)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.90)
- Covered Forest: Fine-grained generalization analysis of graph neural networks (ICML 2025 · similarity 0.90)
- Enhancing Size Generalization in Graph Neural Networks through Disentangled Representation Learning (ICML 2024 · similarity 0.90)
- On the Interplay between Graph Structure and Learning Algorithms in Graph Neural Networks (ICML 2025 · similarity 0.90)
- Towards Bridging Generalization and Expressivity of Graph Neural Networks (ICLR 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.