Enhancing real-time traffic risk prediction with a cost-sensitive learning approach.
TL;DR
This study refines traffic risk classification into four levels, and incorporates misprediction costs into the prediction process through a cost-sensitive learning framework, with the optimal cost coefficients calibrated using a Genetic Algorithm.
Abstract
Real-time traffic risk prediction, enabled by the simultaneous extraction of traffic state variables and their associated risks from vehicle trajectory data, provides a promising approach for proactive traffic safety management. However, existing studies overlook the costs associated with misprediction and the varying consequences of different misprediction types, which undermines the reliability of prediction results. To address these gaps, this study employs empirical data sourced from the NGSIM dataset, from which traffic state variables and risk data aggregated over 5‑second intervals are extracted. Furthermore, this study refines traffic risk classification into four levels, and incorporates misprediction costs into the prediction process through a cost-sensitive learning framework, with the optimal cost coefficients calibrated using a Genetic Algorithm (GA). By integrating this framework with four baseline models, four enhanced models are proposed and systematically evaluated in terms of prediction performance (e.g., precision) and computational efficiency. Results demonstrate that the proposed models consistently outperform their baseline counterparts across multiple evaluation metrics, particularly in identifying high-risk events. Moreover, the computational time of the proposed models remains within acceptable limits for real-time deployment. Reliability analysis further confirms the robustness of the GA-based cost coefficient optimization process.