SimpleST: Efficient Prompt Learning for Traffic Forecasting
A new approach called SimpleST enhances spatio-temporal graph neural networks (GNNs) for traffic prediction by using efficient prompting. The method is lightweight and model-agnostic, allowing pre-trained GNNs to adapt to novel distributions without changing model parameters. This reduces adaptation overhead and complexity, addressing generalization issues caused by spatio-temporal dynamics. The work is detailed in arXiv:2605.08273.
Key facts
- Traffic prediction is essential for transportation systems and urban administration.
- Spatio-temporal GNNs achieve state-of-the-art performance but have low generalization ability.
- Distribution shifts from spatio-temporal dynamics challenge prediction methods.
- SimpleST is a prompt tuning framework for spatio-temporal GNNs.
- It adapts pre-trained GNNs to novel distributions while keeping parameters fixed.
- The prompt mechanism reduces overhead and complexity of adaptation.
- The approach is lightweight and model-agnostic.
- The paper is available on arXiv with ID 2605.08273.
Entities
Institutions
- arXiv