Spatial Analysis of Driving Under the Influence Origins in Seoul Using Geographically Weighted Machine Learning

Driving under the influence (DUI) of alcohol remains a critical public safety concern, accounting for a significant proportion of road traffic injuries and fatalities. Although numerous predictive models have been developed to estimate DUI risk, many conventional approaches such as ordinary least squares (OLS), random forest (RF), and geographically weighted regression (GWR) are constrained by assumptions of spatial homogeneity, linearity, and multicollinearity. To overcome these constraints, this study applies a geographically weighted random forest (GW-RF) model to investigate the spatially varying determinants of DUI incidents in Seoul, South Korea. The analysis targets DUI occurrence density based on trip origin locations within each administrative neighborhood. Explanatory variables include the density of DUI offenses with or without crashes, alcohol-serving land use, transit accessibility, and demographic composition, reflecting environmental exposure and population characteristics. The results indicate that GW-RF achieves competitive predictive accuracy while more effectively accounting for spatial heterogeneity and reducing residual spatial autocorrelation compared with conventional models. Spatial cross-validation further confirms stable out-of-sample generalization across geographically separated partitions. The relative importance and marginal effects of key predictors, including the density of non-crash DUI offenses, alcohol-related crashes, and general food establishments, vary considerably across space. Partial dependence analyses reveal heterogeneous nonlinear effects, including threshold and saturation patterns. A spatiotemporal comparison reveals stronger nighttime clustering and increased influence of alcohol accessibility, indicating time-dependent variation in DUI risk mechanisms. These findings indicate that spatially adaptive machine learning offers an approach for capturing localized risk structures and supports the design of geographically targeted strategies for DUI prevention.