Classification of high-risk paratransit drivers using ensemble machine learning

Abstract In many developing cities, paratransit is the layer of the transport system that absorbs the demand the formal network cannot serve, including during disruptions, which makes the dependability of this informal fleet a practical component of urban transport resilience. That dependability rests almost entirely on individual drivers, yet driver-level risk in the sector remains poorly characterized by conventional statistical approaches. This study develops and validates machine learning classification models to identify high-risk paratransit drivers in Gazipur, Bangladesh. Data were collected through structured face-to-face interviews with 507 active paratransit drivers, conducted at nine major paratransit stations across the district and covering their socioeconomic, behavioral, and operational characteristics as well as self-reported accident history. Five classification algorithms: Logistic Regression, Random Forest, XGBoost, Gradient Boosting, and a Stacking Ensemble were trained and compared to distinguish single-accident from repeated-accident drivers, with performance assessed using repeated stratified cross-validation. Among the models evaluated, Logistic Regression was the best-performing and most stable classifier, and all machine-learning classifiers improved recall of repeated-accident drivers relative to a traditional Poisson-regression benchmark. Feature-importance analysis consistently identified motorized three-wheeler operation as the dominant predictor of repeated-accident status associated with roughly three times the rate of repeated accidents relative to non-motorized modes alongside self-identified driving competence, income, and experience. These findings point to mode-specific regulation of the informal paratransit fleet and provide transportation authorities with an actionable, data-driven tool for targeted safety interventions in informal transport systems.