Driver Allocation Optimization in Ride-Hailing Platforms: A Study Based on Machine Learning Demand Prediction and Min-Cost Max-Flow Modeling

Abstract

Ride-hailing platforms have become an important supplement to urban mobility in cities where public transport coverage is incomplete, transfer convenience is limited, or travel demand fluctuates sharply across time and space. Yet the operational advantages of platform-based mobility do not automatically guarantee efficient matching. In practice, passengers may still face long waiting times, distant drivers, and service shortages in hotspot areas, while other areas simultaneously retain idle supply. Focusing on this mismatch, the present study proposes an integrated forecasting-and-optimization framework for driver allocation in the Bangkok taxi context. Because complete platform order data and real-time driver status records are not publicly available, the study adopts a literature-constrained simulation design. The urban study area is divided into 60 spatial zones and organized into 10-minute time slots. On this basis, an XGBoost model is trained to predict zone-level demand for the next time slot, and the prediction output is then transferred to a min-cost max-flow dispatch model that reallocates limited supply across zones under capacity and distance constraints. All modeling and experimental procedures were implemented in Python and executed on Google Colab. The empirical results show that XGBoost outperforms a lag-based baseline on the test set, with a Test MAE of 2.0920 and a Test RMSE of 2.9884. In the case of sufficient supply, the best solution is to handle the distribution within each region without cross-regional scheduling. However, in the case of peak demand and tight supply, the total demand is predicted to reach 720, while the available supply is only 590. At this time, the min-cost max-flow model can achieve 81.94% demand satisfaction rate, keep 83.73% distribution in the original area, control the proportion of cross-regional distribution at 16.27%, and keep the average distribution distance at 0.2051. These results show that the combination of demand forecasting based on machine learning and interpretable network flow optimization can improve the quality of driver allocation when resources are limited, and provide a practical prototype scheme for the operation improvement of Bangkok and similar cities.