Spatial associations with estimated net migration in Osaka, 2015–2020: A grid-level machine learning approach

As Japan faces rapid depopulation, distinguishing net migration from natural population changes is important for understanding local population dynamics at fine spatial scales. This study proposes a data-driven framework to identify factors associated with census-derived estimates of net migration in Osaka Prefecture at a high-resolution 500 m grid level over the period 2015–2020. Methodologically, we first refined the census survival-ratio method by introducing calibration parameters fitted to prefecture-level Basic Resident Register migration totals, thereby improving age-specific aggregate alignment. Subsequently, spatial associations were analyzed using a machine learning pipeline based on LightGBM. To address the multicollinearity inherent in spatial socioeconomic data—a persistent challenge in urban modeling—we integrated variable clustering with Owen values, a game-theoretic attribution method. This approach allowed us to estimate model-attributed contributions while accounting for groups of highly correlated variables. The analysis revealed “mosaic-like” dynamics characterized by sharp intra-regional contrasts. Specifically, we identified spatial patterns among younger populations that were consistent with the “return to the city” phenomenon during the study period and strongly associated with high-rise condominiums, contrasting with estimated net outflows from aging suburban public housing districts. These findings demonstrate that this high-resolution, explainable AI approach can capture complex, non-uniform spatial patterns of estimated net migration and provide an exploratory basis for identifying area characteristics that warrant further investigation.