XGBackprop
PyTorch implementation of backpropagating through decision trees
Overview
This package implements two flavors of backpropagation through XGBoost decision trees. SHAPBackpropLayer and XGBackpropLayer.
The SHAPBackpropLayer approximates the gradients of a decision tree by Shapley values. the Shapley values of the input features are treated as the gradient df(x) / dx.
The XGBackpropLayer approximates the gradients by relaxing the decision tree branches into sigmoids and samples paths through the decision tree in the style of Straight-Through (ST) approximators for traditional Deep Neural Network (DNN) architectures.