We propose a data-driven optimization model to reduce riders' wait time for vehicle guidance and rebalancing operations, considering the rider demands are under uncertainty. Instead of assuming a pre-defined rider demand distribution, we propose a data-driven framework that integrates Mixture Density Networks (MDNs) and a two-stage stochastic programming model. The integrated framework can compute high-quality guidance and rebalancing solutions that benefit drivers and riders in the ride-hailing system by leveraging the time-series historical data from real data sets. To prove the performance and effectiveness of our approach, we conduct a group of simulations based on the New York High Volume For-Hire Vehicle (HVFHV) trip records. The validation results show that the proposed method outperforms the data-driven deterministic models using GRU and moving average methods. Most significantly, the riders' average wait time using our proposed approach can be reduced by 75.9% compared to the batched matching mechanism.

In this paper, we study a one-to-one matching ride-sharing problem to save the travellers' total travel time considering travel time uncertainty. Unlike the existing work where the uncertainty set is assumed to be known or roughly estimated, in this work, we propose a learning-based robust optimization framework to handle the issue properly. Specifically, we assume the travel time varies in an uncertainty set which is predicted by a machine learning approach- ARIMA using travel time historical data, the predicted uncertainty set then serves as the input parameter for the robust optimization model. To evaluate the proposed approach, we conduct a group of numerical experiments based on New York taxi trip record data sets. The results show that our proposed data-driven robust optimization approach outperforms the robust optimization model with a given uncertainty set in terms of total travel time savings. Further, the proposed approach can improve the travel time savings up to 112.8%, and 34% by average. Most importantly, our proposed approach is capable of handling the uncertainty in a more effective way when the uncertainty degrees become high.