Research map / Graph learning / Knowledge and temporal graphs
STRAP: Spatio-Temporal Pattern Retrieval for Out-of-Distribution Generalization
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 STRAP, a retrieval-augmented STGNN framework that stores spatial, temporal, and spatio-temporal subgraph patterns in key-value libraries and fuses retrieved values into predictions to improve out-of-distribution generalization on streaming graph forecasting.
Problem
STGNNs generalize poorly under spatio-temporal out-of-distribution shifts in streaming graph data, where both temporal dynamics and graph structure change, and existing continual-learning methods suffer forgetting and stability-plasticity trade-offs.
Approach
Build three key-value pattern libraries (spatial, temporal, spatio-temporal) with handcrafted topological and time-series keys and frozen STGNN-derived values, retrieve top-k similar patterns at inference, and fuse them with current representations via a convex combination weight, plugged into any STGNN backbone.
Key results
- STRAP achieves average MAE 20.39 on Air-Stream, 13.31 on PEMS-Stream and 4.85 on Energy-Stream with the STGNN backbone, the best among compared methods in the table.
- The paper reports an average 7.17% improvement across metrics (MAE 4.88%, RMSE 9.12%, MAPE 7.5%) over baselines.
- In few-shot Energy-Stream experiments with 50% missing data on TGCN, STRAP reaches MAE 6.29 and MAPE 78.92%, reported as 8.7% better MAE than the best baseline EWC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Forman-Ricci curvature features (Spatial 3) to the spatial pattern keys (ablation) | STRAP · Spatio-temporal forecasting (spatial pattern analysis) | MAPE (%): performance drop of 6.32%; MAPE of 49.97% when removed | Helped |
| Add statistical and spectral descriptors (Temporal 1) to temporal pattern keys (ablation) | STRAP · Spatio-temporal forecasting (temporal pattern analysis) | MAPE (%): performance degradation of 7.96%; MAPE of 50.61% when removed | Helped |
| Add multi-resolution wavelet analysis (Temporal 2) to temporal pattern keys (ablation) | STRAP · Spatio-temporal forecasting (temporal pattern analysis) | MAPE (%): 44.98% | Helped |
Surprising findings
- Replay achieves the best MAPE on Air-Stream at horizon 3 among baselines and is close to STRAP on some metrics, and the paper notes Retrain is comparable to Pretrain on several settings.
Limitations
- Computational demands grow with larger pattern libraries.
- Relies on availability of robust historical datasets.
- Handcrafted keys and heuristic library management (importance sampling by inverse key norm) lack thorough justification.
Experimental setup
Datasets: AIR-Stream PEMS-Stream ENERGY-Stream
Benchmarks: AIR-Stream PEMS-Stream ENERGY-Stream
Metrics: MAE RMSE MAPE
Baselines: Pretrain Retrain TrafficStream ST-LoRA STKEC EAC EWC Replay ST-Adapter GraphPro PECPM
Builds on: STGCN ASTGCN TrafficStream EWC
Closest papers by meaning
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.87)
- Over-squashing in Spatiotemporal Graph Neural Networks (NeurIPS 2025 · similarity 0.87)
- Efficient Unlearning for Spatio-temporal Graph (Student Abstract) (AAAI 2025 · similarity 0.86)
- State Space Models on Temporal Graphs: A First-Principles Study (NeurIPS 2024 · similarity 0.86)
- Graph Structure Learning for Spatial-Temporal Imputation: Adapting to Node and Feature Scales (AAAI 2025 · similarity 0.86)
- Learning from Highly Sparse Spatio-temporal Data (NeurIPS 2024 · similarity 0.85)
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.