Uncovering traditional retail location competition with machine learning: The role of urban environments in light-asset and capital-intensive formats

Understanding spatial competition between retail formats is vital for sustaining urban commercial vitality. In China’s rapidly evolving urban economy, the contrast between light‑asset and capital‑intensive formats within traditional retail has been increasingly shaped by urban redevelopment and changing consumer behavior. However, prior studies largely confine themselves to linear assumptions, neglecting nonlinear dynamics and complex interdependencies. They also lack systematic comparisons between light-asset and capital-intensive formats, especially when attempting to integrate objective urban features with perceptual dimensions. This study investigates how objective urban environment indicators and human perception jointly shape the spatial density of two traditional retail formats, light‑asset (convenience stores) and capital‑intensive (large-scale supermarkets and shopping malls), in Shenzhen, China. Using multi‑source geospatial data and an integrated framework combining XGBoost, SHAP, Partial Dependence Plots (PDP), Interpretive Structural Modeling (ISM), and Bayesian Networks (BN), we quantify nonlinear effects, interaction patterns, and probabilistic association pathways. Results reveal distinct locational logics: light‑asset retail thrives through widespread penetration, responding strongly to competition intensity, service‑function proximity, and perceptual synergies; capital‑intensive retail concentrates in high‑yield hubs, driven by facility networks, accessibility, and high‑quality aesthetic environments above defined thresholds. Human perception emerges as a core determinant, accounting for up to 28.95% of explanatory importance, often exerting threshold‑dependent effects. The ISM–BN analysis uncovers multi‑entry vs. sequential pathways to spatial clustering across formats. Findings advance retail location theory by linking built‑environment metrics with street-level perceptual attributes, offering actionable guidance for urban planners to tailor format‑specific, perception‑oriented, and threshold‑sensitive strategies for sustainable retail ecosystems.