Research on electric bicycle-motor vehicle conflict severity prediction model for multiple conflict-type scenarios

Electric Bicycle-Motor Vehicle (EB-MV) conflicts pose critical safety challenges at signalized intersections in China. This study proposes a machine learning-based framework for EB-MV conflict prediction by integrating high-resolution aerial trajectory data. Four Conflict Types were classified based on collision angles, and Conflict Risk Levels (minor/moderate/severe) was quantified through surrogate safety measures (SSMs) combining proximity thresholds and kinematic indicators. To address class imbalance, hybrid-sampling is applied before training GBDT and CatBoost prediction models. The results demonstrate the CatBoost model achieves superior performance for Level 1, significantly outperforming GBDT. This work establishes a validated methodological pipeline for intersection conflict analysis, delivers empirical evidence for CatBoost’s superior efficacy in traffic risk prediction, and lays a computational foundation for real-time safety warning systems.