Geographically Weighted Machine Learning

Music City Center 

Spatially-varying models capture relationships between explanatory variables and outcomes that change across a spatial domain, making them essential for addressing spatial problems. In contrast, traditional machine and deep learning models rely on a single, global fit, often failing to account for spatial heterogeneity. To bridge this gap, we introduce a spatially-weighted decorrelation transformation, inspired by local and geographically weighted regression principles. This approach enables machine and deep learning models to be applied locally at each spatial location while maintaining a smooth, continuous transition across the domain. The result is a spatially-adaptive machine learning framework that enhances predictive performance in the presence of spatial non-stationarity.