Predictability and interpretability of hybrid link-level crash frequency models for urban arterials compared to cluster-based and general negative binomial regression models

Machine learning (ML) techniques have higher prediction accuracy compared to conventional statistical methods for crash frequency modelling. However, their black-box nature limits the interpretability. The objective of this research is to combine both ML and statistical methods to develop hybrid link-level crash frequency models with high predictability and interpretability. For this purpose, M5′ model trees method (M5′) is introduced and applied to classify the crash data and then calibrate a model for each homogenous class. The data for 1134 and 345 randomly selected links on urban arterials in the city of Charlotte, North Carolina was used to develop and validate models, respectively. The outputs from the hybrid approach are compared with the outputs from cluster-based negative binomial regression (NBR) and general NBR models. Findings indicate that M5' has high predictability and is very reliable to interpret the role of different attributes on crash frequency compared to other developed models.