Statistical Spatial Interpretable Machine Learning in R Using Tree Ensembles and SHAP Values

This handbook chapter aims to present, discuss, and explore the uses of statistical machine learning algorithms and interpretable machine learning tools in the context of spatial analysis. In this regard, the chapter is mostly aimed towards researchers and practitioners in urban and regional spatial analysis, and the field of regional science in general. Using a relatively simple dataset downloadable as part of an R package, the chapter applies a series tree-based machine learning models – with XGBoost being the primary one, and analyzes the results using SHAP values. The use of spatial features, spatial cross-validation, and spatial dependence are focal topics. The use of coordinates, spatially lagged features, and their consequences on predictions are investigated by taking into account potential data leakage caused by proximities over space of data instances in calibration and validation sets. The chapter demonstrates the advantages of the used techniques for spatial analysis while highlighting the possible drawbacks of internalizing spatial information into machine learning models. In doing so, models predicting urban noise levels are employed.