Research map / Graph learning / GNN architectures
Decoupling the Depth and Scope of 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 |
|---|---|---|---|
| decouple depth and scope: apply GNN on a PPR subgraph instead of full L-hop neighborhood | 3-layer GCN · Yelp | F1-micro: 0.4028 -> 0.5291 | Helped |
| decouple depth and scope: PPR subgraph for GAT | GAT-SAINT · ogbn-arxiv | accuracy: 0.6977 -> 0.7235 | Helped |
| use ShaDow with PPR scope instead of GraphSAINT | GCN-SAINT 3-layer · ogbn-products | accuracy: 0.8003 -> 0.7742 | Hurt |
| increase ShaDow depth from 3 to 5 without expanding scope (ablation) | ShaDow-GCN (PPR) · ogbn-products | accuracy: 0.7742 -> 0.7821 | Helped |
| add subgraph pooling (ablation) | ShaDow-GCN (PPR) · ogbn-products | accuracy: 0.7821 -> 0.7914 | Helped |
| use PPR sampler instead of 2-hop sampler (ablation) | ShaDow-SAGE · ogbn-products | accuracy: 0.7935 -> 0.7905 | Hurt |
| use PPR sampler instead of 2-hop sampler (ablation) | ShaDow-SAGE · Yelp | F1-micro: 0.6542 -> 0.6538 | No clear effect |
| use ShaDow-SAGE on link prediction vs top leaderboard | SEAL · ogbl-collab | Hits@50: 0.5371 -> 0.5492 | Helped |
| use ShaDow-GAT with scope 400 vs 200 (ablation) | ShaDow-GAT · ogbn-papers100M | test accuracy: 0.6681 -> 0.6708 | Helped |
| extract local subgraph for SGC as power increases to 40 (ablation) | SGC · oversmoothing test | accuracy | Helped |
Closest papers by meaning
- Mixture of Scope Experts at Test: Generalizing Deeper Graph Neural Networks with Shallow Variants (NeurIPS 2025 · similarity 0.91)
- Graph Neural Networks Do Not Always Oversmooth (NeurIPS 2024 · similarity 0.89)
- A Flexible, Equivariant Framework for Subgraph GNNs via Graph Products and Graph Coarsening (NeurIPS 2024 · similarity 0.89)
- Towards Scalable and Deep Graph Neural Networks via Noise Masking (AAAI 2025 · similarity 0.89)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.89)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · 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.