Analyzing spatial heterogeneity of ridesourcing usage determinants using explainable machine learning

Abstract

There is a pressing need to study spatial heterogeneity of ridesourcing usage determinants to develop better-targeted transportation and land use policies. This study incorporates spatial information (i.e., the geographic coordinates of census tracts) into the machine learning model and leverages state-of-the-art explainable machine learning techniques to analyze census-tract-to-census-tract ridesourcing usage, identify the key factors that shape the usage, and explore their nonlinear associations across different spatial contexts. Specifically, we analyze the spatial heterogeneity of ridesourcing travel in Chicago based on three spatial contexts, including downtown, neighborhood and airport. The results reveal that built environment variables collectively contribute to the largest importance for the downtown and airport context, while socioeconomic and demographic variables are the strongest predictors for the neighborhood context. Travel cost, the number of commuters and transit supply variables have evident nonlinear associations with ridesourcing usage, and these associations show strong differences across these three spatial contexts. Moreover, incorporating geographic coordinates is shown to be useful in improving model's capability to capture spatial information and thus enhance its predictive performance. These findings provide transportation professionals with location-based insights to better plan and manage ridesourcing services in Chicago.