Introducing Spatial Heterogeneity via Regionalization Methods in Machine Learning Models for Geographical Prediction: A Spatially Conscious Paradigm

This study addresses the challenge of incorporating spatial heterogeneity in predictive modeling by introducing regionalization methods in the preprocessing step of the modeling workflow. Spatial heterogeneity, where the mean of attribute values varies across spatial units, poses difficulties for traditional models. To tackle this, we propose a novel approach called Regionalization Random Forest (RegRF), which combines Random Forest with regionalization techniques to enhance predictive performance. Regionalization combines multiple spatial objects into homogeneous regions, which are incorporated into predictive models, allowing models to capture local variations. This research investigates three key questions: (1) How does the predictive performance of RegRF vary when constructed using different regionalization techniques? (2) How does RegRF compare to benchmark methods, including both spatial statistical approaches and spatially conscious machine learning models like Geographically Weighted Random Forest (GW-RF)? Five regionalization methods—WARD, AZP, Kmeans, SKATER, and Max-p—are tested on datasets of varying sizes. Results show that RegRF significantly improves performance over "non-spatial" Random Forest models with minimal additional computation time. While RegRF performs competitively with Geographically Weighted Regression, it requires much less computational effort. GW-RF was not outperformed on smaller datasets but failed to complete for larger datasets. These findings suggest that RegRF can enhance machine learning models by accounting for spatial phenomena, with potential for further optimization.