When Space Matters, How Can We Explain Random Forest?

SIRUS is a new stable rule extraction algorithm for regression and classification problems designed for explainability purposes. The general principle of SIRUS is to extract rules from Random Forests (RF). This algorithm inherits a level of accuracy comparable to RF and state-of-the-art rule algorithms producing much more stable and shorter lists of rules. In this work, we extend SIRUS for the case of spatially correlated data in a regression problem. In particular, we propose to combine SIRUS with the RF-GLS algorithm instead of the classical RF in order to make the estimation procedure spatially aware. A simulation study, based on pseudo-real data, will be used to assess how the spatial correlation in the data affects the explainability capability of the proposed algorithm.