Publications

Proposes LGPBoost, a unified STAT-ML framework that couples tree-boosting with GP mixed effects to jointly model nonlinear covariate effects and spatio-temporal (road-segment and annual) dependencies, validated via simulation and Florida motorcycle crashes (2014–2023).

This study investigates the impact of autonomous vehicles on intercity public transport mode choice within the Beijing-Tianjin-Hebei urban agglomeration using a hybrid random parameter logit model, revealing how multi-layered heterogeneity and travel distance influence traveler preferences.

This study utilizes a partially temporally constrained random-parameters logit model to investigate spatiotemporal disparities in motorcycle injury severity in Florida (2020-2022), integrating macro-level socio-demographic factors with micro-level crash data.

This study proposes a Multiscale Geographical Random Forest (MGRF) integrating street-view semantic visual features for traffic safety modeling by capturing multiscale spatial heterogeneity.

This study jointly analyzes intercity travelers’ adjustments to departure dates and travel modes during adverse weather using a correlated random parameters bivariate probit model, revealing significant correlations and heterogeneity in adaptive behaviors.

This study utilizes interpretable machine learning (XGBoost with SHAP) to unravel the non-linear effects of socio-demographic and commute patterns on spatial inequalities in overall and fatal traffic crash rates, demonstrating superior performance over traditional spatial statistical models.

This study utilizes spatial lag models and geographically weighted regression to quantify spatial inequities in traffic injury rates in New York City (2021-2023), integrating urban road network topology and social vulnerability indices.

Using joint random-parameters bivariate probit models and a partially temporal constrained framework, this paper quantifies how helmet use and crash/context factors jointly shape moped riders’ injury severity in Florida (2019–2021), and reveals COVID-19-related temporal instability in key effects.

Proposes an SST-DHL framework that combines dynamic hypergraph learning with self-supervised representation learning to improve sparse, skewed urban traffic accident forecasting and enhance interpretability.

Using CRSS intersection crash data (2016–2018), this study models serious-injury risk as a two-dimensional, non-parametric function of two-vehicle speeds. The framework moves beyond conventional univariable impact-speed definitions, revealing crash-type-specific and non-monotonic speed–risk patterns that inform intersection safety interventions.

This study proposes an RPBPHM framework to jointly model truck- and car-driver injury severities in the same crash using UK STATS19 (2017–2019), quantifying cross-equation correlation while capturing multilayer unobserved heterogeneity and temporal stability/instability of contributing factors.