Taxi Trip Travel Time Prediction with Isolated XGBoost Regression

Travel time prediction is crucial in developing mobility on demand systems and traveller information systems. Precise estimation of travel time supports the decision-making process for riders and drivers who use such systems. In this paper, static travel time for taxi trip trajectories is predicted by applying isolated XGBoost regression models to a set of identified inlier and extreme-conditioned trips and the results are compared with other existing best models in this context. XGBoost uses an ensemble of decision trees and is robust to outliers and thus it is believed to perform well on time series predictions. We show that, compared to other existing best models, XGB-IN (XGBoost prediction model of in-lier trips) model prediction values reduce mean absolute error as well as root mean squared error and exhibit impressive correlation with actual travel time values while XGB-Extreme model is able to provide reasonably accurate prediction results for a set of extreme-conditioned trips with shorter actual time durations. We demonstrate the achievability of travel time prediction with XGBoost regression and show that our approach is applicable to large-scale data and performs well in predicting static travel time.