Spatially aware gradient boosting towards sparser models

Lars Knieper, University of Goettingen

Co-authors: Hannah Miller, University of Goettingen; Elisabeth Bergherr, University of Goettingen

Abstract: One of the key features of model-based component-wise gradient boosting is a data-driven variable selection mechanism which takes place while estimating the effects simultaneously. When spatial effects, of areal or point-reference data, are included as a potential model-component a drastic increase of chosen fixed effects can be observed. This is accompanied by a high selection frequency of the spatial component without achieving an impactful reduction of the loss function.

To address this ineffective variable selection, we propose to eliminate the competition between fixed and spatial effects by separating the spatial part from the component-wise mechanism and ensuring complete estimation in each iteration. Additionally, we suggest using spatial cross-validation, which accounts for the autocorrelated structure of spatial data when constructing folds.