Explainable machine learning for identifying nonlinear associations between street vitality and the built environment in a Historic District

Historic districts exhibit fine-grained spatial variation, while street vitality may relate nonlinearly to the built environment. Using 213 micro-sampling points in the Lihuangpi Road Historic District, Wuhan, this study integrates heritage location, functional facilities, transport accessibility, street networks, frontage characteristics, and streetscape visual features. OLS, GAM, and XGBoost were compared, while SHAP, LOWESS, and spatial validation were used to assess nonlinear associations, model interpretation, and reliability. Under nested random validation, XGBoost achieved R² values of 0.474, 0.490, and 0.460 for overall, weekday, and weekend vitality, respectively, outperforming GAM and OLS. However, spatial GroupKFold reduced these values to 0.012, 0.084, and 0.052, with further deterioration as training–test separation increased. In the original ten-predictor models, Distance Heritage Core ranked first (29.7%–30.0%); after adding X and Y coordinates and distance to the study-area boundary, its importance decreased to 10.6%–11.3% and its rank fell to third or fourth, indicating that part of its apparent contribution reflected general spatial position. The results show that explainable machine learning can reveal nonlinear within-area associations, but random-validation performance should not be interpreted as spatial transferability. The framework supports micro-scale diagnosis and cautious, heritage-compatible renewal in historic districts.