Geographically weighted machine learning for modeling spatial heterogeneity in traffic crash frequency and determinants in US

Spatial analyses of traffic crashes have drawn much interest due to the nature of the spatial dependence and spatial heterogeneity in the crash data. This study makes the best of Geographically Weighted Random Forest (GW-RF) model to explore the local associations between crash frequency and various influencing factors in the US, including road network attributes, socio-economic characteristics, and land use factors collected from multiple data sources. Special emphasis is put on modeling the spatial heterogeneity in the effects of a factor on crash frequency in different geographical areas in a data-driven way. The GW-RF model outperforms global models (e.g. Random Forest) and conventional geographically weighted regression, demonstrating superior predictive accuracy and elucidating spatial variations. The GW-RF model reveals spatial distinctions in the effects of certain factors on crash frequency. For example, the importance of intersection density varies significantly across regions, with high significance in the southern and northeastern areas. Low-grade road density emerges as influential in specific cities. The findings highlight the significance of different factors in influencing crash frequency across zones. Road network factors, particularly intersection density, exhibit high importance universally, while socioeconomic variables demonstrate moderate effects. Interestingly, land use variables show relatively lower importance. The outcomes could help to allocate resources and implement tailored interventions to reduce the likelihood of crashes.