Prediction models for understanding traffic safety risk on rural roads in Serbia
摘要
This paper presents a traffic accident prediction model developed for IB-category rural roads in the Republic of Serbia, utilizing the largest dataset analyzed in the country to date. The model employs interpretable machine learning techniques to assess the impact of road geometry, traffic exposure, and quality on accident frequency. The model exhibits robust predictive accuracy with a R2 of 79%, especially in pinpointing high-risk regions linked to increased traffic density and geometric limitations. SHAP-cluster analysis was used to interpret variable contributions and reveal hidden patterns in cluster-specific risk profiles, which can be used to define and prioritize prevention measures. Future study will expand this framework by extending it to various road types, including freeways, and by investigating additional elements, such as user behavior and weather conditions, that may improve forecast accuracy. The results provide a practical tool for road authorities and policymakers aiming to improve traffic safety on the rural road network.