A multi-source, multi-stage framework for ride-hailing safety management.

The rapid growth of ride-hailing services has transformed urban mobility, but the predominance of non-professional drivers raises pressing safety concerns. Existing research often lacks a comprehensive framework that integrates trajectory, environmental, and driver-related data to thoroughly examine risky driving behaviors. To address this gap, this research proposes a framework to analyze, predict, and intervene in speeding and jerky driving among ride-hailing drivers. It also explores the underlying factors contributing to such behaviors and offers targeted improvement strategies. The analysis uses a rich multi-source dataset, including over six million GPS trajectories, OpenStreetMap road network, and real-time weather data, along with 450 self-reported questionnaires from ride-hailing drivers. A multi-stage framework is employed, consisting of three components: (i) a bivariate random parameter model with heterogeneity in means to identify contributing factors to risky driving behavior; (ii) a Dual-Task Gated Cross-Attention Network for real-time risk prediction and proactive safety monitoring; (iii) a Gradient Boosting Classifier model to investigate factors influencing drivers’ willingness to reduce risky behaviors. The analysis illustrates that, while speeding and jerky driving share common risk factors, such as trip distance and acceleration, they diverge substantially in their sensitivity to infrastructure and geometric conditions. The questionnaire survey analysis reveals several factors that influence a driver’s willingness to improve their behavior: driving experience, driver ratings, risk perception, income level, and crash history. Overall, the proposed framework offers a robust, data-driven approach for managing risky driving in ride-hailing services, supporting real-time supervision and targeted policy development to enhance urban road safety.