Abstract: This study forecasts municipal-level population change in Japan using machine learning. We employ total population from the Basic Resident Register and regional well-being indicators published by SCI Japan as features, and model spatial dependence through a Graph Convolutional Network (GCN) constructed from adjacency relationships among administrative units. The experimental design comprises (i) a baseline model, (ii) incorporation of missing-value imputation, (iii) k-nearest neighbor (kNN) neighborhood-size optimization, and (iv) prefecture-level Leave-One-Group-Out Cross-Validation (LOGO-CV). Primary evaluation metrics are MAE, RMSE, and $$R^{2}$$. Results show MAE consistently near 1%, with accuracy improving when imputation and spatial dependence are incorporated; performance peaks at approximately k = 22. Under LOGO-CV, $$R^{2}$$ tends to be negative in prefectures with extremely small variance, warranting cautious interpretation. Preprocessing includes standardization and prefecture-level median imputation. Distributional analysis indicates that population growth rate g concentrates near zero, with greater dispersion among small municipalities. The contributions of this work are: establishing an empirical benchmark for short-horizon small-area population change; quantifying the effects of spatial dependence and imputation; and presenting a reproducible evaluation protocol based on LOGO-CV.

Keywords: Small-Area Population Forecasting, Graph Convolutional Networks, WellBeing Indicators, Spatial Dependence, Cross-Validation

Pages: 216-224