Zero loss training and cv improvements

Hi folks
I was reading this method ( and thinking what could be done with xgboost, but i think we need some changes to make it work

The absolute operator should be used in a “mini batch” instead of each row

It’s a incredible regularization method and i thin it could improve our out of sample resulta here

Anyone can help? I can contact some of researchers from this paper if needed

Does the algorithm require mini-batch training? Currently, XGBoost uses the entire training data to perform gradient boosting.