Probabilistic Prediction of Trip Travel Time and Its Variability Using Hierarchical Bayesian Learning

AbstractThis paper proposes a probabilistic machine learning methodology to predict travel time and its variability for trips between locations in New York City. First, a hierarchical Bayesian generalized linear regression model was trained to estimate predictive distribution of the trip’s unit-distance travel time conditional on trip features. This intermediate step isolates the effect of trip distance to capture the impact of traffic conditions and other covariates on trip travel time. Then, trip travel time and its variability were obtained given the predictive distribution of normalized travel time and the trip’s fastest path distance. Specifically, a Bayesian regression model with varying coefficients was fitted to capture the effect of temporal variations of traffic conditions and spatial effect of geographic regions on model parameters and in turn on trip travel time. The model was trained using New York City ambulance and taxi trip data and its performance was verified with benchmarks methods. Although it shows promising performance, the proposed methodological framework estimates the posterior distributions of the model’s interpretable parameters and eventually estimates the predictive distributions for trip travel times. The estimated distributions are required for uncertainty quantification to make more reliable decisions in traffic management and traffic control strategies and achieve proactive enhancement in the transportation systems. The proposed approach emphasizes the potential use of origin-destination trip data for travel time prediction, quantifying its spatiotemporal variations, and capturing the predictive uncertainties, which are central to operational transportation systems’ design and performance evaluation.