Explainable AI for Urban Retail Site Selection: SHAP-PDP-Bayesian Network Modeling of Facility Synergy and Perceptual Thresholds
Main article
Abstract
Urban retail site selection increasingly depends on complex interactions among consumer mobility, facility complementarity, commercial competition, street-level perception, and transport accessibility. Conventional retail-location models are often useful for descriptive mapping but provide limited support for explaining nonlinear thresholds and conditional facility synergies. This study develops an explainable artificial intelligence framework for evaluating urban retail site suitability with multi-source geospatial data. The framework integrates point-of-interest density, mobility proxies, road-network accessibility, neighboring facility functions, and computer-vision-derived streetscape perception into a 500 m grid representation. XGBoost is used to estimate nonlinear suitability patterns for convenience-format and anchor-format retail sites, SHAP values identify the relative contribution and directional role of each determinant, partial dependence plots reveal threshold effects, and Bayesian network modeling converts model explanations into probabilistic decision rules. The empirical demonstration shows that the XGBoost model improves predictive performance over linear and semi-parametric baselines. Facility synergy and perceptual quality jointly explain more than one-third of model importance for anchor-format sites, while competition intensity and service-gap signals are more decisive for convenience-format sites. The analysis further identifies several decision-relevant thresholds, including minimum facility-density and perception-quality levels beyond which site suitability increases sharply. The results indicate that explainable AI can support site screening, commercial land-use planning, and retail network optimization without reducing urban retail decisions to opaque prediction scores.
