Open lopuhin opened 7 years ago
A related ticket is https://github.com/TeamHG-Memex/eli5/issues/56; the description was unclear, but what I meant is something opposite of this ticket: provide an option to show features for all targets in a single combined table, without splitting them by target, similar to how feature importances work now.
For trees and tree ensembles that perform multiclass classification or multitarget regression and have separate trees for different targets it's maybe more informative to view feature importances per target. Maybe this should be optional, on by default for such kinds of models. Related sklearn issue: https://github.com/scikit-learn/scikit-learn/issues/8055 For xgboost, feature importances are parsed from get_dump, so it's easy to separate trees for different targets.