Although Classification and Regression Trees (CART) are widely used for prediction, they do not exploit spatial information and may therefore underperform in geo-referenced settings characterized by spatial effects. We propose the Spatially Aware Regression Tree (SPART), an extension of CART for continuous responses that incorporates spatial information into the splitting rule through a penalized spatial criterion. The method discourages non-contiguous partitions, thereby accommodating spatial instabilities and improve accuracy. A preliminary application to the Boston housing dataset suggests that SPART outperforms CART in spatial prediction tasks.
Machine learning and the prediction of geographically distributed data: A methodological proposal.
Cartone A
;Mazzaferro S;Postiglione P
2026-01-01
Abstract
Although Classification and Regression Trees (CART) are widely used for prediction, they do not exploit spatial information and may therefore underperform in geo-referenced settings characterized by spatial effects. We propose the Spatially Aware Regression Tree (SPART), an extension of CART for continuous responses that incorporates spatial information into the splitting rule through a penalized spatial criterion. The method discourages non-contiguous partitions, thereby accommodating spatial instabilities and improve accuracy. A preliminary application to the Boston housing dataset suggests that SPART outperforms CART in spatial prediction tasks.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
