Bayesian Optimized XGBoost Model for Traffic Speed Prediction Incorporating Weather Effects

Accurate traffic speed prediction is essential to devise traffic control strategies, travelling plans, identifying congestion, reducing travel times and related intelligent decision making. Large amounts of historical traffic speed and weather data contain complex non-linear interdependencies, but at the same time, incorporating weather effects in short to medium term travel speed prediction has been much less explored in the existing literature. With the growth in the amount of traffic speed-related data, tree-boosting algorithms like XGBoost have empirically proven to be very efficient for predictive modeling. However, this algorithm necessitates several hyperparameters to be optimized and searched in a whole complex parameter space. Moreover, using the traditional grid and random search is computationally expensive and an inefficient way in the context of traffic Big data sets. To address these challenges we propose an approach using a Bayesian-based Optimization for systematic exploration of the complex parameter space and including of weather conditions variables. Experiments were conducted using travel speed data collected through speed detectors and weather information for Manhattan, New York. Extensive data preprocessing, missing value imputation, feature selection using sequential feature selection and Bayesian optimization were implemented. The results demonstrated that the proposed approach is able to extract complex patterns from multivariate data comprising speed and weather variables lags for a more accurate prediction, particularly for medium-range horizon.