Geographical-XGBoost: a new ensemble model for spatially local regression based on gradient-boosted trees

XGBoost is a widely used machine learning method known for exhibiting high accuracies in regression and classification tasks. Current efforts to create spatial versions of XGBoost do not alter the original XGBoost algorithm; nor do they implement any geographical modification. Here, we fill these gaps by introducing Geographical-XGBoost (G-XGBoost), which extends XGBoost in multiple ways. First, G-XGBoost creates local models by assigning spatial weights linked directly to the model’s accuracy by focusing on the samples that cause the most error. Second, it applies ensemble architecture, by combining the global and local models, for training, validation, and prediction, which improves model accuracy. Third, G-XGBoost calculates local feature importance using spatial weights, something not proposed so far. We evaluate G-XGBoost on six benchmark datasets against three global regression models (Ordinary Least Squares, Random Forests, and XGBoost) and three spatially local regression models (Geographically Weighted Regression, Geographical Random Forests, and Geographically Weighted Random Forests). G-XGBoost outperforms all existing models improving R2 by 17.44% and mean absolute error by 17.42%, on bootstrapped mean values. G-XGBoost is also an exploratory tool as it analyzes how feature importance varies in space, contributing thus to the need for interpretable and explainable AI, and therefore, it can further advance spatial machine learning application to geographical studies.