Comparative Analysis of Linear Regression, Spatial Models, and Machine Learning Models in Exploring the Relationship between the Built Environment and Bike-Sharing Demand

Abstract

Understanding the relationship between the built environment and bike-sharing demand is crucial for developing effective transport policies and promoting sustainable urban mobility. Numerous analytical models have been applied to this problem, including generalized linear regression (GLR), spatial econometric models, and machine learning (ML) methods. However, differences in model assumptions and learning mechanisms often yield divergent findings. Relying solely on a single modeling approach therefore risks introducing bias into interpretation and may lead to ineffective policy decisions. Using Shenzhen, China, as a case study, this research investigates the built environmental determinants of bike-sharing demand through negative binomial regression (NBR), geographically weighted regression, random forest, and gradient boosted decision trees. We develop a comprehensive workflow to compare, select, and implement these modeling approaches and to translate model outputs into planning-relevant evidence on the relative importance of environmental factors, the presence of spatial heterogeneity, and the emergence of nonlinear, threshold, or saturation effects. Our results show partial consistency in identifying key environmental factors and general agreement in the direction of associations. GLR and spatial models are particularly useful for early-stage planning and exploratory analysis with limited data, providing clear identification of significant variables and overall directionality. ML models, when supported by larger data sets, reveal nuanced nonlinear patterns and thresholds that are valuable for designing more targeted policy interventions. The study underscores that model selection is not merely a technical choice but a critical decision that shapes the evidence base for bike-sharing planning and built environment interventions.