Metropolis-scale resilient and trustworthy traffic flow inference using multi-source data
The Task-Aware Attentive Neural Process (TA-ANP) predicts traffic flows and quantifies both aleatoric and epistemic uncertainty by fusing floating car data with sparse fixed-detector records. The model was evaluated on a newly built dataset covering 26,166 road segments with 2,371 sensor-equipped segments, successfully performing real-time estimation and forecasting at unobserved locations while adapting to altered sensor setups without retraining.
Why it matters — The framework allows traffic monitoring systems to maintain high accuracy and calibrated uncertainty during sensor failures or reconfigurations without requiring model retraining. It also provides reliable uncertainty metrics that reduce the number of physical sensors needed for optimal coverage.