Research map / Graph learning / GNN architectures
S$^3$GNN: Efficient Global Mixing and Local Message Passing for Long-Range Graph Learning
Dai Shi, Luke Thompson, Linhan Luo, Lequan Lin, Andi Han, Junbin Gao, José Miguel Hernández-Lobato
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add spatial propagation and linear feature transform to low-rank global mixing (S3GNN vs diagonal filter) (ablation) | diagonal spectral filtering dynamic (Eq. 7) · peptides-struct | MAE: 0.6824 -> 0.2429 | Improved |
| replace S2GNN spectral dynamics with S3GNN | S2GNN · Graph Property Prediction Eccentricity | log10(MSE): -1.4622 -> -2.4801 | Improved |
| replace S2GNN with S3GNN (no eigendecomposition, no PE) | S2GNN · peptides-func | AP: 72.75 -> 73.20 | Improved |
| replace S2GNN with S3GNN | S2GNN · peptides-struct | MAE: 0.2487 -> 0.2429 | No clear effect |
| replace Stable-ChebNet with S3GNN | Stable-ChebNet · ogbn-arxiv | accuracy: 75.42 -> 76.38 | Improved |
| replace Stable-ChebNet with S3GNN on Barbell graph (N=50) | Stable-ChebNet · Barbell graph MSE | MSE: 0.17 -> 0.005 | Improved |
| use S3GNN in the reasoning GNN of KGQA | Reasoning GNN · WebQSP multi-hop (no RAG) | F1: 58.8 -> 69.6 | Improved |
| use S3GNN in reasoning GNN with RAG | Reasoning GNN-RAG · CWQ multi-entity | F1: 64.8 -> 64.9 | No clear effect |
| replace GCN backbone with S3GNN in topological Schrodinger bridge matching | TSB-BM · Brain signals | 2-Wasserstein distance: 7.51 -> 7.25 | Improved |
| replace processor blocks of MeshGraphNet with S3GNN layers | MGN · Cylinder Flow | min test loss: 0.173 -> 0.102 | Improved |
| use antisymmetric constraint on W versus free W (ablation) | S3GNN · Barbell graph | MSE | Improved |
| use orthogonal constraint on W instead of antisymmetric (ablation) | S3GNN · Barbell graph | MSE | Hurt |
Closest papers by meaning
- Return of ChebNet: Understanding and Improving an Overlooked GNN on Long Range Tasks (NeurIPS 2025 · similarity 0.88)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.87)
- Adaptive Message Passing: A General Framework to Mitigate Oversmoothing, Oversquashing, and Underreaching (ICML 2025 · similarity 0.87)
- On Oversquashing in Graph Neural Networks Through the Lens of Dynamical Systems (AAAI 2025 · similarity 0.87)
- Accurate and Scalable Graph Neural Networks via Message Invariance (ICLR 2025 · similarity 0.87)
- Beyond Oversquashing: Understanding Signal Propagation in GNNs Via Observables (NeurIPS 2026 · similarity 0.87)
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.
