Predicting Road Accident Risk

Abstract: Road accidents constitute one of the leading causes of fatalities and economic losses worldwide. Predicting the likelihood of such incidents using data-driven approaches can significantly enhance road safety management and resource allocation. This paper presents an ensemble-based machine learning framework for predicting road accident risk, integrating multiple gradient boosting models—XGBoost, LightGBM, and CatBoost. The proposed ensemble combines the predictive strengths of each model through weighted averaging to minimize Root Mean Square Error (RMSE) and improve generalization across diverse driving conditions. Extensive experiments were conducted on the Kaggle Playground Series (Season 5, Episode 10) dataset, which contains multi-dimensional traffic, environmental, and temporal attributes. The ensemble achieved an RMSE of 0.1346, outperforming individual learners and demonstrating the effectiveness of hybrid boosting in accident risk assessment. The study provides valuable insights into the influence of key features such as speed limit, road surface condition, and weather index, offering a scalable model for intelligent transport and safety analytics.

Keywords: Road Accident Risk, Machine Learning, Ensemble Learning, Gradient Boosting, Traffic Prediction.