Thanks for the PR. While reviewing #17541 I did some experiment with a synthetic imbalanced dataset with RFs or nonlinear preproc + logistic regression, or even just logreg, and each time it seemed that class_weight="balanced" would either change nothing or actually hurt the performance, both for threshold-insensitive metrics (e.g. ROC AUC and average precision), log loss, Brier score or threshold sensitive metrics balanced accuracy, f1 score, matthrews corrcoef and others.

So while I am not opposed to implement this for scikit-learn's HGBDT, I think we should better warn the users that class_weight="balanced" is not necessarily a good idea.

Furthermore, there exist a weird pitfall with class_weight="balanced" documented in #10233 although I can probably not affect HGBDT as they are likely to be underfitting to the point of just learning an constant intercept.