Interpretable Machine Learning for Crash Severity Analysis of Food Delivery Motorcyclists I Gede Brawiswa Putra*, Febrian Fitryanik Susanta, Bimo Harya Tedjo, Pei-Fen Kuo
Department of Geomatics, National Cheng Kung University, Taiwan, *10903014[at]gs.ncku.edu.tw
Abstract
The COVID-19 pandemic has significantly increased the demand for online food delivery services. This surge has intensified competition among platforms and placed greater pressure on delivery riders to prioritize speed over safety. As a result, crash incidents involving food delivery motorcycles have nearly doubled compared to those used for routine commuting. While previous studies have focused on general motorcycle crashes, few have examined the specific factors influencing crash severity among delivery riders. Additionally, most existing research relies on traditional spatial models, which may fail to capture nonlinear relationships and spatial heterogeneity. Another challenge is the imbalance in severity data, with serious and fatal crashes underrepresented, limiting the reliability of standard statistical analysis.
To address these gaps, this study applies GeoShapley, an explainable machine learning (XAI) framework that captures both spatial and non-spatial effects. GeoShapley treats geographic location as an interactive predictor, allowing for interpretable, location-specific insights. We also apply the Synthetic Minority Oversampling Technique (SMOTE) to improve class balance and reduce model bias.
The analysis uses crash data from 2,314 food delivery motorcycle incidents recorded in Taipei City in 2020. Results show that severe crashes are more likely on roads with higher speed limits, straight segments, intersections, and in suburban areas near restaurants. Male riders and signal violations are also strongly linked to higher crash severity. GeoShapley reveals that these risk factors vary significantly across locations, highlighting the importance of spatial heterogeneity in crash modeling. This study demonstrates the benefits of combining interpretable machine learning with spatial analysis and class-balancing techniques. The findings offer practical insights for developing targeted, location-specific safety interventions for food delivery motorcyclists in urban areas.
