-
XGBoost pitfalls: the mistakes that quietly wreck a boosted model
A boosted-tree model can look healthy while being wrong in ways ordinary metrics do not expose. It may overfit after the useful trees were built, leak the test set into tuning, scramble columns at inf