AXIL: Exact Instance Attribution for Gradient Boosting

arXiv:2301.01864v2 Announce Type: replace-cross Abstract: We derive an exact, prediction-specific instance-attribution method for fitted gradient boosting machines (GBMs) trained with squared-error loss, with the learned tree structure held fixed. Each prediction can be written as a weighted sum of training targets, with coefficients determined only by the fitted tree structure and learning rate. These coefficients are exact instance attributions, or AXIL weights. Our main algorithmic contributi