Research map / Label-efficient and robust learning / Recognition and long-tail
FlashST: A Simple and Universal Prompt-Tuning Framework for Traffic Prediction
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.
FlashST is a spatio-temporal prompt-tuning framework that pre-trains a traffic forecasting backbone on PEMS datasets, then freezes it and tunes a lightweight prompt network with an InfoNCE-based uniformity loss to adapt to unseen downstream traffic datasets.
Problem
Traffic prediction models generalize poorly under distribution shift when applied to new downstream datasets, and end-to-end or full fine-tuning is costly and transfers knowledge poorly.
Approach
A lightweight spatio-temporal prompt network (temporal and Laplacian-eigenvector spatial context plus temporal/spatial dependency encoders) is trained for in-context adaptation while the downstream model is frozen, with a uniformity (InfoNCE-style) loss on prompt embeddings to map pre-training and downstream data into a shared distribution.
Key results
- On PEMS07(M), FlashST achieves MAE 2.59 / RMSE 5.25 / MAPE 6.54%, the best among 13 end-to-end baselines (best baseline PDFormer MAE 2.68).
- On CA-D5, FlashST reports MAE 13.26 and RMSE 21.83, versus MTGNN 13.90 / 22.44 and PDFormer 13.85 / 22.59.
- Plugging FlashST into MTGNN gives MAE 2.31 on ChengDu-DIDI, and the paper states it reduces baseline training time by 20% to 80%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FlashST prompt-tuning to downstream predictors | spatio-temporal prediction models · PEMS07(M) | MAE: 2.67 -> 2.59 | Helped |
| Add FlashST prompt-tuning to downstream predictors | spatio-temporal prediction models · CA-D5 | MAE: 13.63 -> 13.26 | Helped |
| Add FlashST prompt-tuning to downstream predictors | spatio-temporal prediction models · ChengDu-DIDI | MAE: 2.33 -> 2.31 | Helped |
| Add FlashST prompt-tuning to downstream predictors | spatio-temporal prediction models · NYC Citi Bike | MAPE: 48.69% -> 49.77% | Hurt |
| Add temporal context distillation (ablation) | FlashST with MTGNN · PEMS07(M) and CA-D5 | prediction error | Helped |
| Add spatial context distillation (ablation) | FlashST with MTGNN · PEMS07(M) and CA-D5 | prediction error | Helped |
| Add temporal dependency encoder (ablation) | FlashST with MTGNN · PEMS07(M) and CA-D5 | prediction performance | Helped |
| Add spatial dependency encoder (ablation) | FlashST with MTGNN · PEMS07(M) and CA-D5 | prediction performance | Helped |
| Add unified distribution mapping (ablation) | FlashST with MTGNN · PEMS07(M) and CA-D5 | prediction performance | Helped |
| Replace unified distribution mapping with batch normalization (ablation) | FlashST with MTGNN · PEMS07(M) and CA-D5 | prediction performance | Hurt |
| Vary the temperature coefficient and loss weight coefficient (ablation) | FlashST · PEMS07(M) and CA-D5 | prediction performance | No clear effect |
Surprising findings
- Direct full fine-tuning of pre-trained baselines often gives only marginal gains over end-to-end training, and sometimes hurts (e.g., STGCN w/ Finetune on PEMS07(M) MAE 3.18 vs 3.14 end-to-end).
- Pre-trained models without adaptation (w/o Finetune) perform very poorly on target datasets, e.g., GWN CA-D5 MAE 27.28 vs 13.63 end-to-end.
Limitations
- Prompt-tuning results rely on re-pre-training for each ablation variant and on a fixed set of pre-training PEMS datasets.
- Some FlashST gains are small (e.g., MAPE on NYC Citi Bike is worse than MTGNN, 49.77% vs 48.69%).
- Only traffic-related datasets are evaluated; generalization to other spatio-temporal domains is untested.
Experimental setup
Datasets: PEMS03 PEMS04 PEMS07 PEMS08 PEMS07(M) CA-D5 ChengDu-DIDI NYC Citi Bike
Benchmarks: PEMS07(M) CA-D5 ChengDu-DIDI NYC Citi Bike
Metrics: MAE RMSE MAPE
Baselines: TGCN STGCN ASTGCN GWN STSGCN AGCRN MTGNN STFGNN STGODE DMSTGCN MSDR STWA
Builds on: infoNCE graph convolution prompt-tuning
Closest papers by meaning
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.87)
- Decomposed Spatio-Temporal Mamba for Long-Term Traffic Prediction (AAAI 2025 · similarity 0.86)
- Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting (AAAI 2025 · similarity 0.86)
- Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph Forecasting (ICLR 2025 · similarity 0.86)
- TESTAM: A Time-Enhanced Spatio-Temporal Attention Model with Mixture of Experts (ICLR 2024 · similarity 0.85)
- Spatio-Temporal Few-Shot Learning via Diffusive Neural Network Generation (ICLR 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.