Three Common Machine Learning Algorithms Neither Enhance Prediction Accuracy Nor Reduce Spatial Autocorrelation in Residuals: An Analysis of Twenty-five Socioeconomic Data Sets

Machine learning (ML) is being applied in an increasing volume of geographical research. However, the aspects of spatial autocorrelation (SAC) in the residuals produced by ML models have been understudied compared to the benefit of ML, namely, reduction of prediction errors. In this study, we examined the relationship between predictive accuracy and the reduction in the residual SAC for 597 variables from 25 geographical socio-economic data sets using spatial and nonspatial cross-validation of three ML algorithms such as random forests, support vector machine, and artificial neural network (ANN) to provide an extensive empirical diagnosis—but not a definitive theory—of the relationship between SAC and ML. Our results highlighted that the ML algorithms with tuned hyperparameters yielded marginal predictive accuracy gains and the minimal decreases in residual SAC. ANN revealed lower accuracy and higher reduction in the residual SAC than others. This implies ML algorithms in geographical research in socio-economic domains would not always result in higher prediction accuracy. We suggest that ML in geographical research should be cautiously employed when the main objective is related to the residual SAC. We also showed that spatial cross-validation neither improves predictive accuracy substantially nor reduce the residual SAC effectively.