Exploring effects of expressway accident on drivers’ adaptive behavior in different scenarios: A comparative evaluation of machine learning techniques and discrete choice models

The congestion and time loss caused by expressway accidents are significant, and existing studies have not fully explored the differences and mechanisms of drivers' adaptive behaviors before, and during expressway driving. Based on 30,000 decision choices from 2500 drivers in western Japan, this study uses latent class analysis (LCA), XGBoost (eXtreme Gradient Boosting), and discrete choice model (DCM) to identify key influencing factors, compare model performance, and make policy recommendations. The LCA divided drivers into three categories; Results of XGBoost shows that there are significant differences in the importance of factors such as age, trip purpose, and accident clearance time among different scenarios and driver groups, and the model outperformed DCMs in the individual cluster analysis, with cluster 3 achieving a prediction accuracy of 61.1 % (area under the curve value); and SHAP (SHapley Additive Explanation) analysis indicates that information accuracy, alternative expressways and queue dynamics are dominative factors affecting Cluster 1, 2 and 3 respectively by using important features, and young drivers are more sensitive to real-time information than middle-aged and older drivers, and commuters are significantly more inclined to change routes than leisure travelers (P < 0.01). It is recommended to strengthen visual early warning for low-sensitivity groups, provide accident details for safety-priority groups, optimize the layout of expressway entrances and exits, and implement hierarchical information release based on age and travel purpose. This study provides an empirical basis for the design and policy formulation of dynamic transportation systems across scenarios and clusters.