Stacking models for analyzing traffic injury severity on two-lane, two-way rural roads.

Abstract

The analysis of injury severity in accidents allows traffic management agencies to assess crash risk more effectively and develop cost-effective interventions. The aim of this research is to present a two-layer stacking model as a means of forecasting accident severity. In the initial layer, the system incorporates benefits derived from many base classification algorithms through a three-stage process to evaluate the outcomes of each model configuration. These base algorithms include Random Forests, Decision Tree, K Nearest Neighborhood and Support Vector Machine; in the second layer, Logistic Regression and Random Forest algorithms are used to classify crash injury severity. In total, 24,141 traffic accidents were recorded on 135 two-way, two-lane roads. The process of model calibration entails the optimization of several parameters, such as the number of trees in three fundamental methods of classification, the learning rate and the regularization coefficient which is achieved by the utilization of a systematic grid search strategy. To validate the model, the Stacking model's performance is assessed in comparison to other conventional models. The results indicate that the Stacking model has greater performance. Consequently, each component included in the prediction of severity is categorized into distinct groups according to its impact on results.