Using Bayesian Optimized XGBoost for Predicting Traffic Accident Duration
TL;DR
The proposed model incorporates temporal, weather, environmental, and infrastructure features from the traffic accident dataset, and XGBoost hyperparameters are tuned using Bayesian optimization, and outperforms conventional machine learning algorithms in both recall and F1 score.
Abstract
Traffic accidents adversely affect socioeconomic conditions and human well being by causing casualties, property damage, congestion, and secondary crashes. Accurate prediction of incident duration is critical for effective TIM and optimal resource allocation. Traditional studies on accident duration prediction have mainly used statistical models, but because of the explosion in data volume and the recent progress in machine learning, data-driven approaches have now become dominant. Therefore, this paper proposes a hybrid model that combines Bayesian Optimization (BO) with Extreme Gradient Boosting (XGBoost) for predicting traffic accident duration. The proposed model incorporates temporal, weather, environmental, and infrastructure features from the traffic accident dataset, and XGBoost hyperparameters are tuned using Bayesian optimization. As a result, the BO-XGBoost model outperforms conventional machine learning algorithms in both recall and F1 score, thus demonstrating its effectiveness for predicting accident duration. More importantly, the feature analysis clearly shows that temporal factors can have a major impact on accident duration prediction. Therefore, this model serves as an excellent, interpretable decision-support tool for traffic management authorities to allocate resources optimally and ease traffic congestion.