Explainable artificial intelligence visions on incident duration using eXtreme Gradient Boosting and SHapley Additive exPlanations

Abstract

Efficient management of traffic incidents is a focal point in traffic management, with direct implications for road safety, congestion, and the environment. Traditional models have grappled with the unpredictability inherent in traffic incidents, often failing to capture the multifaceted influences on incident durations. This study introduces an application of Explainable Artificial Intelligence (XAI) using eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) to analyze the complexities of traffic incident duration prediction. Utilizing a substantial dataset of over 366,000 records from the Houston traffic management center, the study innovates in the domain of traffic analytics by predicting incident durations and revealing the contribution of each predictive variable. The XGBoost algorithm's ability to handle multi-dimensional datasets was employed to identify crucial variables affecting incident durations. Meanwhile, SHAP values offered transparency into the model's decision-making process, clarifying the roles of over fifty parameters. The study's results demonstrate that variables such as the involvement of heavy trucks and blockage of main lanes are essential in influencing incident durations, aligning with findings from previous literature. The SHAP analysis further revealed time-sensitive patterns, with time of day and day of the week exhibiting considerable effects on predictions. The beeswarm plots of SHAP provided a detailed visualization of these effects, differentiating between high and low values effects for each variable. The model's high accuracy, with a coefficient of determination (R²) of 0.72 and a root mean square error (RMSE) of 21.2 min, indicates the potential of XAI in enhancing traffic management systems.