Cold-Start Traffic State Forecasting at Unseen Sensor Locations via Support-Conditioned Meta-Graph Learning

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Traffic sensors provide real-time data, but traffic management requires future predictions.
• Expanding sensor networks presents a challenge for pre-trained forecasting models due to new, unseen locations.
• Existing spatio-temporal graph neural networks struggle with sensor-specific embeddings for new network nodes.

Purpose of the Study:

• To develop a method for adapting pre-trained traffic forecasting models to previously unseen sensor locations.
• To address the 'cold-start' problem in network-level traffic prediction during sensor network expansion.
• To enable accurate traffic speed or flow forecasts at newly added sensor locations.

Main Methods:

• Proposed support-conditioned sensor-adaptive meta-graph learning (SC-SAMG) framework.
• Utilized a support-set encoder, task-specific graph learner, and meta-learning.
• Adapted network-level forecasters using 1-7 days of target observations.
• Employed a leakage-controlled held-out-node protocol for evaluation.

Main Results:

• SC-SAMG consistently outperformed fine-tuned GRU, AGCRN, and DCRNN baselines.
• Achieved up to 11% relative reduction in Mean Absolute Error (MAE) compared to the adaptive-graph baseline.
• Demonstrated up to 7% relative MAE reduction compared to the diffusion convolutional baseline.
• Successfully forecasted traffic speed or flow for the next 15-60 minutes at new sensor locations.

Conclusions:

• Support-conditioned graph adaptation is a viable approach for incorporating new sensor locations into traffic forecasting.
• SC-SAMG effectively solves the cold-start problem for adaptive-graph forecasters.
• The proposed method enhances the scalability and applicability of network-level traffic prediction systems.