Single-vehicle crash severity outcome prediction and determinant extraction using tree-based and other non-parametric models

Single-vehicle crashes are more fatality-concentrated and have posed increasing challenges in traffic safety, which is of great research necessity. Tremendous previous studies have conducted relevant analysis with econometric modeling approaches, whereas the ability of non-parametric methods to predict crash severity is still smattering of knowledge. Consequently, the main objective of this paper is to conduct single-vehicle crash severity prediction with different tree-based and non-parameter models. An alternate aim is to identify the intrinsic mechanism of how contributing factors determine single-vehicle crash severity. By virtue of Grid-Search method, this paper conducted fine-tuning of different models to obtain the best performances based on five crash severity sub-datasets. For model evaluation, the accuracy indicators were calculated in training, validation and test sets, respectively. Besides, feature importance extraction was undertaken based on the results of model comparison. The finding indicated that these models didn't exhibit a huge performance difference for crash severity prediction in the same severity level; however, the performances of the models did vary among different datasets, with an average training accuracy of 99.27 %, 96.4 %, 86.98 %, 86.84 %, 71.76 % in fatal injury, severe injury, visible injury, complaint of pain, PDO crash datasets, respectively. Additionally, it was found that in each severity dataset, the indicator urban freeways is a determinant factor that leads to the occurrence of crashes while rural freeways is more related to more severe crashes (i.e., fatal and severe crashes). This paper can provide valuable information for model selection and tuning in accident severity prediction. Future research could consider the influences that temporal instability of contributing features has on the model performances.