Balancing Temporal Prediction Errors in Urban Mobility Demand Modeling Using a Hybrid Loss Function

Accurate taxi demand prediction plays a critical role in efficient urban mobility management, supporting fleet allocation, congestion mitigation, and service reliability. However, conventional loss functions such as mean squared error (MSE) are inherently scale-dependent and tend to emphasize high-demand periods, as larger absolute errors dominate the optimization process. This characteristic often results in disproportionately large relative errors during periods of low demand, where accurate prediction remains essential for operational efficiency and cost control. To address this limitation, this study introduces a hybrid MSE–MAPE loss function that jointly considers absolute and relative prediction errors, enabling more balanced learning across heterogeneous demand levels in spatiotemporal taxi demand modeling. The proposed loss function is evaluated using New York City taxi data aggregated at an hourly resolution. Experimental results demonstrate that the hybrid loss significantly reduces mean absolute percentage error (MAPE) across different time periods, achieving improvements of up to ۳۱%, while incurring only a marginal increase in RMSE. These findings suggest that incorporating relative-error information into the training objective provides a simple yet effective mechanism for improving robustness and overall predictive balance in taxi demand prediction.

کلیدواژه ها:

Taxi Demand Modeling ، Hybrid Loss Function ، Spatiotemporal Prediction ، Deep Learning