• check_classification_targets should warn the user if the number of unique classes is larger than say 50% of n_samples: the user is likely feeding bad targets which can be highly problematic when the size of the model depends linearly (or more) on the number of classes (e.g. Gradient Boosting models, OneVsOne SVM, ...).

• check_classification_targets which calls type_of_target triggers a call to _assert_all_finite and np.unique(y) which are redundant with checks done elsewhere (e.g. in check_X_y)

• check_classification_targets should probably be refactored to return the array of unique classes (to be used for the classes_ attribute of the classifier) to avoid redundant computation.

• check_classification_targets should probably be used more consistently for any classifier maybe by being used internally in check_X_y or in _validate_data instead of being called manually on a case by case basis.

And we probably need a common test to check the above (e.g. all classifier should raise the warning).