Ensemble-based machine learning approach to prioritize driving safety indicators using connected vehicle system data for proactive safety analytics

Driving safety indicators can quantitatively estimate crash potential, which enables proactive evaluations of traffic safety. Real-time traffic safety evaluation can contribute to crash prevention by identifying hazardous sections immediately. To implement an effective traffic safety evaluation model, selecting indicators that have a significant impact on identifying the presence of hazardous situations is necessary. The purpose of this study was to design a methodology for prioritizing driving safety indicators for effective safety evaluation. A set of driving safety indicator candidates were extracted from 74-km and 91-day vehicle trajectory data collected by connected vehicle systems on a Korean freeway. Hazardous and normal sections were categorized using the crash data for each segment. Support vector machine (SVM), artificial neural network (ANN), and K-nearest neighbor (KNN) models were constructed by utilizing driving safety indicators as independent variables and setting hazardous and normal sections as dependent variables. Then, those three models were ensembled using a voting method. The proposed ensemble model achieved a maximum accuracy 4.7% higher than those of the individual models, which implied that combining multiple models could improve the reliability and robustness of the results. In addition, the prioritization of 24 driving safety indicators was derived by summarizing the variable importance results from each model using permutation importance, a model-independent feature importance method that can derive feature importance with a consistent computational process regardless of the model. The outcome of this study is expected to be utilized as a valuable foundation for selecting indicators in safety evaluations for proactive traffic safety management.