Research map / Label-efficient and robust learning / Domain adaptation and generalization
Drawing Informative Gradients from Sources: A One-stage Transfer Learning Framework for Cross-city Spatiotemporal Forecasting
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 TSJT, a one-stage joint training framework for cross-city few-shot traffic forecasting that filters source-city gradients (Target-Skewed Backward) and adds a Node Prompting Module as an add-on to STGCN and Graph WaveNet.
Problem
Spatiotemporal forecasting in data-scarce target cities; pre-trained transfer models absorb noise and harmful knowledge from data-rich source cities due to distribution shift.
Approach
Train directly on source and target data jointly in one stage, but keep only the components of source-city gradients that do not increase target loss (projecting out the part opposing the target gradient and dropping the parallel-opposing part), and use a prompt bank of shared spatiotemporal patterns to make nodes with similar patterns receive similar gradients.
Key results
- TSJT with GWN achieves about 4.14% improvement over the best baseline STGP on average across datasets.
- TSJT (GWN) averages a 13.47% improvement over standard GWN trained with Reptile.
- Removing NPM increases MAE by roughly 20% on the Shenzhen dataset in ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TSJT with GWN instead of STGP | Cross-city few-shot forecasting · METR-LA | MAE (avg.): 3.58 -> 3.38 | Helped |
| Use TSJT with GWN instead of STGP | Cross-city few-shot forecasting · PEMS-BAY | MAE (avg.): 2.19 -> 2.07 | Helped |
| Use TSJT with GWN instead of STGP | Cross-city few-shot forecasting · Didi-Chengdu | MAE (avg.): 2.42 -> 2.30 | Helped |
| Use TSJT with GWN instead of STGP | Cross-city few-shot forecasting · Didi-Shenzhen | MAE (avg.): 2.17 -> 2.15 | Helped |
| Remove the Node Prompting Module (NPM) (ablation) | TSJT (STGCN) · Didi-Shenzhen | MAE: approximately 20% increase | Hurt |
| Remove Target-Skewed Backward (TSB) (ablation) | TSJT (STGCN) · Didi-Chengdu and Didi-Shenzhen | forecasting accuracy | Hurt |
| Remove the knowledge graph A_K (ablation) | TSJT (STGCN) · Didi-Chengdu and Didi-Shenzhen | forecasting accuracy | No clear effect |
| Increase the node-prompt bank size from a small size to a sufficient size (ablation) | TSJT · PEMS-BAY | MAE and RMSE | Helped |
| Use fewer source cities (ablation) | TSJT, GPD, and STGP · Didi-Chengdu and METR-LA | MAE | Mixed |
Surprising findings
- The knowledge graph term A_K has a negligible effect on performance in ablation.
Limitations
- Evaluated only on traffic speed/flow datasets, with the authors noting it is taken as an example task.
- Uses only relatively simple backbones (STGCN, GWN); TSJT(STGCN) trails STGP on some settings.
- Larger prompt bank raises time and memory cost.
Experimental setup
Datasets: PEMS-BAY METR-LA Didi-Chengdu Didi-Shenzhen
Benchmarks: METR-LA PEMS-BAY Didi-Chengdu Didi-Shenzhen
Metrics: MAE RMSE
Baselines: HA ARIMA DCRNN GWN DSTAGNN FOGS AdaRNN ST-GFSL DASTNet TPB TransGTR GPD
Builds on: Reptile ST-GFSL TPB GPD
Closest papers by meaning
- Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework (NeurIPS 2024 · similarity 0.87)
- GeoPro-Net: Learning Interpretable Spatiotemporal Prediction Models Through Statistically-Guided Geo-Prototyping (AAAI 2025 · similarity 0.87)
- Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach (NeurIPS 2025 · similarity 0.87)
- Diffusion Transformers as Open-World Spatiotemporal Foundation Models (NeurIPS 2025 · similarity 0.86)
- Spatio-Temporal Few-Shot Learning via Diffusive Neural Network Generation (ICLR 2024 · similarity 0.86)
- How Different from the Past? Spatio-Temporal Time Series Forecasting with Self-Supervised Deviation Learning (NeurIPS 2025 · similarity 0.86)
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.