A Comparative Study Using Generalized Ordered Probit, Stacking Ensemble, and TabNet: Application to Determinants of Pedestrian Crash Severity

In this study, we introduce the adaptive interpretive modeling (AIM) pipeline, utilizing the TabNet model to address some deep learning (DL) challenges within traffic safety analysis. The AIM pipeline is specifically designed to tackle issues such as causal inference, interpretability, class imbalance, decision-making transparency, and generalizability of findings. To evaluate the efficacy of our AIM pipeline, which includes the TabNet model, we conducted a comparative analysis with two models: the generalized ordered probit (a statistical model) and the stacking ensemble model, which is based on machine learning (ML) techniques. This comparison was aimed at ascertaining the relative performance of these models in crash severity analysis. For the application of these methodologies, we utilized pedestrian crash data from Utah, covering the years 2010–2021. The results from this comparative study reveal that the AIM pipeline, anchored by the TabNet model, not only achieves higher prediction accuracy compared to both the generalized ordered probit and the stacking ensemble model but also enhances model interpretability, ensuring the robustness and generalizability of the findings. This study contributes to the traffic safety literature by evaluating the efficacy of advanced machine learning and deep learning techniques in analyzing pedestrian crash severity, offering insights that could inform future safety measures and interventions.