I’m studying the implementation of XGBoost and the concepts about it, so I read in some places “margin”. But I don’t understand what is the concept about this margin, likewise I didn’t find it in the paper with this name. I would be glad if anyone explain me it better. Thanks.
The base_margin is comparable to passing another booster as the base for a training process.
I encountered this when you have a gradient that has no definition for the starting point, you train with another loss function and use the result as the base_margin for your real boosting round