ABSTRACT Machine learning is increasingly applied in spatial data analysis, with spatial features offering potential improvements in its predictive accuracy. However, comparative evaluations of different spatial feature types remain limited, especially under varying relationships between input and response variables and differing levels of spatial autocorrelation. This study systematically evaluates the effectiveness of three spatial feature types, including spatial coordinates, spatial lags, and Moran eigenvector spatial filters (MESF), across four machine learning models with distinct estimation procedures. Using simulation experiments, the analysis examines both first‐order effects (correlation between input and response variables) and second‐order effects (spatial autocorrelation in the response variable). To validate the practical applicability of the simulation findings, an empirical analysis is subsequently conducted using municipal‐level homeownership rate data from South Korea. Results from both the simulation and empirical analysis show that spatial features improve model accuracy more when first‐order effects are weak. MESF consistently provides the greatest accuracy improvement, primarily due to its capacity to capture diverse spatial patterns through orthogonal eigenvectors. Under strong second‐order effects, spatial features again enhance accuracy, with MESF proving most effective. MESF also best reduces residual spatial autocorrelation by modeling spatial dependence without confounding input‐response relationships, whereas spatial lags often overcorrect, reversing autocorrelation direction. These findings highlight that the effectiveness of spatial features varies more by feature type than by model type. By quantifying their impact on accuracy and residual autocorrelation, this study provides empirical support for incorporating spatial features in machine learning, offering foundational insights for spatial machine learning development.
Koo et al. (Fri,) studied this question.