Recent years have witnessed tremendous growth in mobile application-based ride-hailing taxi services, primarily due to the convenience, flexibility, and affordability associated with such services. A key challenge faced by these fast-growing taxi services is the imbalance between taxi demand and driver supply. In this thesis, we investigate three crucial aspects that can influence demand predictions and demand-supply equilibrium: (i) spatial tessellation, (ii) spatio-temporal modeling, and (iii) anomaly detection.



By evaluating two popular tessellation schemes on diverse demand data sets, we find that the performance of each strategy is dependent on the city geography, spatial distribution of the data, and the time of the day, and that neither strategy can perform optimally across the entire forecast horizon. To that end, we propose an online ensemble learning algorithm to pick the best tessellation strategy at each time step. Next, we perform an extensive comparison of the classic time-series models and recently developed deep learning models, and propose the adoption of a graph-based deep learning model that can learn from spatial partitions represented as nodes on arbitrarily structured graphs. Finally, we develop efficient deep learning-based solutions based on extreme value theory for anomaly detection in transportation networks.