Research map / Graph learning / Knowledge and temporal graphs
Robust Spatio-Temporal Centralized Interaction for OOD 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 STOP, a spatio-temporal OOD model that replaces node-to-node messaging with centralized low-rank attention through learnable context units, perturbs this messaging with learnable masks, and trains with a spatio-temporal DRO objective.
Problem
Spatio-temporal graph neural networks relying on node-to-node messaging degrade under spatio-temporal distribution shifts and structural changes (node additions and removals), and generalize poorly to inductive new nodes.
Approach
Replace node-to-node message passing with centralized interaction through a small set of Context-Aware Units using multi-head low-rank attention, perturb the aggregation with learnable masking units (GenPU), and optimize a worst-case-environment DRO objective over the generated environments.
Key results
- STOP achieves up to 17.01% improvement in OOD generalization performance over 14 baselines across six datasets.
- On SD, STOP reaches MAE 17.71 at horizon 3 versus 18.44 for the next best (STONE).
- STOP reports up to 18.44% improvement in inductive learning performance; on SD inductive MAE is 22.74 versus 25.06 for STONE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STOP instead of the strongest listed baseline | Spatio-temporal prediction · LargeST-SD, horizon 3 | MAE: 18.44 -> 17.71 | Helped |
| Use STOP instead of the strongest listed baseline | Spatio-temporal prediction · LargeST-GBA, horizon 3 | MAE: 20.19 -> 18.33 | Helped |
| Use STOP instead of the strongest listed baseline | Spatio-temporal prediction · PEMSD3-Stream, horizon 3 | MAE: 12.79 -> 11.39 | Helped |
| Use STOP instead of the strongest listed baseline | Spatio-temporal prediction · KnowAir, horizon 6 | MAE: 26.19 -> 24.37 | Helped |
| Use STOP instead of the strongest listed baseline | Inductive learning on new nodes · LargeST-SD | MAE: 25.06 -> 22.74 | Helped |
| Use STOP instead of the strongest listed baseline | Inductive learning on new nodes · LargeST-GBA | MAE: 26.28 -> 24.76 | Helped |
| Remove node-to-node graph messaging (ablation) | STGCN · LargeST-SD, horizon 3 | MAE: 25.72 -> 25.45 | Helped |
| Replace node-to-node interaction with centralized interaction (ablation) | STGCN · LargeST-SD, horizon 3 | MAE: 25.72 -> 24.87 | Helped |
| Remove node-to-node graph messaging (ablation) | STGCN · KnowAir, horizon 6 | MAE: 29.49 -> 26.18 | Helped |
| Replace node-to-node interaction with centralized interaction (ablation) | STGCN · KnowAir, horizon 6 | MAE: 29.49 -> 25.44 | Helped |
| Remove node-to-node graph messaging (ablation) | STAEformer · LargeST-SD, horizon 3 | MAE: 26.20 -> 25.80 | Helped |
| Replace node-to-node interaction with centralized interaction (ablation) | STAEformer · LargeST-SD, horizon 3 | MAE: 26.20 -> 24.65 | Helped |
Surprising findings
- Removing node-to-node messaging improved several advanced STGNNs (e.g., STGCN and STAEformer variants without graph/attention messaging performed better under OOD settings).
Limitations
- Structural shift evaluation is simulated by removing 10% of nodes and adding 30% new nodes, not real-world graph evolution.
- Some baselines had to be modified (removing node embeddings or adaptive graph learning) to fit the ST-OOD setting.
Experimental setup
Datasets: LargeST-SD LargeST-GBA LargeST-GLA LargeST-CA PEMSD3-Stream KnowAir
Benchmarks: LargeST-SD LargeST-GBA LargeST-GLA LargeST-CA PEMSD3-Stream KnowAir
Metrics: MAE RMSE MAPE
Baselines: STGCN GWNet STNorm STID STAEformer STNN D2STGNN BigST RPMixer CaST STONE TrafficStream
Builds on: STID STAEformer Distributionally robust optimization STONE Non-stationary Transformers
Closest papers by meaning
- Over-squashing in Spatiotemporal Graph Neural Networks (NeurIPS 2025 · similarity 0.86)
- Learning from Highly Sparse Spatio-temporal Data (NeurIPS 2024 · similarity 0.84)
- Improving Generalization of Dynamic Graph Learning via Environment Prompt (NeurIPS 2024 · similarity 0.84)
- Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework (NeurIPS 2024 · similarity 0.84)
- Harnessing Feature Resonance under Arbitrary Target Alignment for Out-of-Distribution Node Detection (NeurIPS 2025 · similarity 0.84)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.84)
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.