Road infrastructure has significant effects on road traffic safety and needs further examination. In terms of traffic crash prediction, recent studies have started to develop deep learning classification algorithms. However, given the uncertainty of traffic crashes, predicting the traffic risk potential of different road sections remains a challenge. To bridge this knowledge gap, this study investigated a real-world expressway and collected its traffic crash data between 2013 and 2020. Then, according to the time-spatial density ratio (Pts), road sections were assigned into three classes corresponding to low, medium, and high risk levels of traffic. Next, different classifiers were compared that were trained using the transformed and resampled feature data to construct a traffic crash risk prediction model. Last, but not least, partial dependence plots (PDPs) were employed to interpret the results and analyze the importance of individual features describing the geometry, pavement, structure, and weather conditions. The results showed that a variety of data balancing algorithms improved the performance of the classifiers, the ensemble classifier superseded the others in terms of the performance metrics, and the combined SMOTEENN and random forest algorithms improved the classification accuracy the most. In the future, the proposed traffic crash risk prediction method will be tested in more road maintenance and design safety assessment scenarios.
Keywords: classifiers; feature importance; performance evaluation measures; resampling algorithms; traffic crash risk prediction.