详情

原文传递 Equitable Traffic Crash Prediction Framework To Support Safety Improvement Grants Allocation

题名：	Equitable Traffic Crash Prediction Framework To Support Safety Improvement Grants Allocation
正文语种：	eng
作者：	ZIHANG WEI;ZIHAO LI;MIHIR MANDAR KULKARNI;XIMIN YUE
作者单位：	Zachry Department of Civil and Environmental Engineering at Texas A&M University;Zachry Department of Civil and Environmental Engineering at Texas A&M University;Zachry Department of Civil and Environmental Engineering at Texas A&M University;Zachry Department of Civil and Environmental Engineering at Texas A&M University
摘要：	Traffic crash prediction model is a vital supportive tool for governors to allocate safety improvement grants. The traditional method directly uses the number of severe crashes in previous year as reference to allocate next year's safety improvement grants. However, this method cannot accurately represent the number of crashes in next year due to the time varying nature of crash occurrences. Al-based crash prediction models can improve the overall prediction performance while, the issue of equality in Al-based crash prediction models has been neglected. Applying model with biases for the allocation of safety improvement grants may exacerbate social inequality. Therefore, in order to facilitate safety grants allocation, this study purposes an equitable framework for predicting the number of severe crashes happened in next year by incorporating oversampling technique with Al-based models. The HSIS crash data and sociodemographic data from North Carolina are utilized as a study case. Specially, this study applies XGBoost model and TabNet model to improve the overall model performance. However, albeit the improved performance, these models increase the model inequality. In order to alleviate the inequality issue, this study applies the SMOTE to balance the training dataset, thereby reducing dataset bias. The results show that for both XGBoost and TabNet learning models, the SMOTE can improve the model equality. Moreover, the proposed framework of TabNet learning model with SMOTE is proven to improve the overall prediction performance as well as the model equality within three sensitive group pairs including low-income/high-income, rural/urban, and aging/non-aging census tracts.
出版年：	2023
期刊名称：	ITE journal
卷：	93
期：	9
页码：	37-45