Lateral shifting behavior of vehicles at horizontal curves and its influencing factors: application of LightGBM and SHAP

Horizontal curves are disproportionately associated with severe crashes due to increased vehicle instability and lane departure risks. This issue is particularly critical in low- and middle-income countries (LMICs), where poor lane discipline, mixed traffic, and geometric inconsistencies amplify crash potential. Despite its significance, lateral shifting (LS) behavior on horizontal curves remains understudied in LMIC contexts. This study addresses the research question: What are the key factors influencing the lateral shifting behavior of vehicles on rural horizontal curves, and how can they be modeled and interpreted effectively in LMIC conditions? Using trajectory data from 8,748 vehicles across 18 curve segments in India, an explainable machine learning framework is developed. Light Gradient Boosting Machine (LightGBM) was selected for its superior performance in classification metrics compared to other ML models. SHapley Additive exPlanations (SHAP) were integrated to interpret model outputs and quantify feature contributions, while Shannon entropy was applied to assess prediction uncertainty. Findings reveal that lane type, vehicle speed, curve radius, lateral clearance, superelevation, and traffic interactions, especially oncoming vehicles, significantly influence lateral shifting behavior. SHAP analysis uncovers nonlinear effects and interaction patterns, including a threshold response to speed and clearance. Notably, the influence of preceding vehicles differs from oncoming traffic, suggesting asymmetric behavioral responses rarely captured in prior studies. This research fills four major gaps in existing literature related to context, terrain, feature scope, and methodology. It provides data-driven insights to support safer curve design and lane departure countermeasures tailored to LMIC road environments.