Neural nets are state of the art on image and textual data. However, tree boosting algorithms reign supreme when tabular data is considered. In this project we suggest a novel technique of training a Student network which learns from a Teacher XGBoost model, mimicking SHAP values as a surrogate to the rich feature representation usually mimicked when both Student and Teacher models are neural networks.
Project Paper: https://github.com/DanaCohen95/copycat/blob/master/Copycat.pdf