Uncovering the residential relocation during urban renewal with time-series big data and spatial interpretable machine learning

Identifying the spatial pattern and correlates of demolition-triggered residential relocation induced by urban village renewal is crucial for revealing the socio-spatial patterns of urban renewal and optimizing inclusive urban renewal resettlement strategies. While existing studies have built a solid foundation, this work focuses on underexamined yet critical angles: micro-scale single-village patterns, origin-linked place association effects, and multi-scale sensitivity tests against MAUP. Taking Guangzhou, a representative Chinese city for urban village and renewal research, as the case, this study develops a Geospatial eXplainable Artificial Intelligence (GeoXAI)-based framework, using station-level time-series mobile signaling data for relocation identification, the semi-variogram-based procedure and sensitivity analysis to mitigate the Modifiable Areal Unit Problem (MAUP) in data aggregation, and the Geographically Weighted Random Forest (GWRF) model for quantifying nonlinear, threshold and spatially heterogeneous patterns. Results reveal that most relocated residents present a polycentric agglomeration pattern in urban village areas within 10 km of the original village, associated with accessible informal housing with affordable low costs, familiar living contexts, convenient E-bike accessibility, and concentrated local industrial clusters matching their original occupational livelihoods. The complex variable-relocation correlations are fully interpreted via Partial Dependence Plots (PDPs) and distribution of SHapley Additive exPlanations (SHAP). This study enhances relocation research accuracy and interpretability by addressing the MAUP and integrating GeoXAI tools, offering methodological references and evidence-based support for human-oriented urban renewal planning in megacities.