Prediction of high-risk areas using the interpretable machine learning: Based on each determinant for the severity of pedestrian crashes

Abstract

Despite the steady decline in the total number of pedestrian crashes in Korea, the pedestrian fatality rate per 100,000 people remains high compared to the Organization for Economic Cooperation and Development (OECD) average. As the data of traffic crashes is gradually accumulated every year, various machine learning methodologies are needed to analyze this data. This study proposed a new algorithmic approach using Local Interpretable Model-Agnostic Explanation (LIME) to identify vulnerable pedestrian crash areas based on each determinant influencing these severity in Seoul. Using the pedestrian crash data from 2016 to 2018, this study uses the XGBoost to model the determinants of pedestrian crash severity and LIME to predict high-risk areas for each determinant. A new algorithmic approach using LIME was proposed to enhance the reliability by filtering data based on an Explanation Fit (R² ≥ 0.26), in reference to Cohen (1988). Upon synthesizing the results, Cheongnyangni Station and Gangnam Station in Seoul were predicted as vulnerable to severe pedestrian crashes due to the superposition of influencing variables considered in this study. In this study, the heatmap predictions derived from the proposed algorithm methodology provided insights into the vulnerable areas and non-linear determinants of pedestrian crash severity. Additionally, this study suggests policy implications aimed at reducing pedestrian crash severity and enhancing pedestrian safety.