Spatiotemporal Ensemble Modeling for Urban Taxi Travel Time Prediction

Abstract

In this paper, we propose to address the challenge in predicting urban taxi travel time through spatiotemporal ensemble modelling techniques. Accurate travel time prediction is very important not just for planning urban mobility but also for running ridesharing platforms with excellent user experience. We compare a stacking model to a ridge meta-regressor optimised through gradient descent, KNN regressor tuned by hill-climbing, decision tree and random forest.The stacking model emerged as the best, underscoring again the efficacy of ensemble methods in coping with the spatiotemporal complexity of the data. KNN regressors, with their hyperparameters optimised through hill climbing yielded competitive results.Our experimental results show that the ensemble approach, especially the stacking model, generally has much higher performance in prediction than individual models for taxi travel time. This proves that the ensemble approach can effectively work as an alternative way to predict real-time urban transportation system states.