Discussion: is get_features_name_out something missing or on purpose?

[MarieSacksick]

Hello :) !

I am trying to access the features names outside at the end of a pipeline where I use a GapEncoder and a DropCol, because I need them to plot feature importance with feature names beside.

My question is: is it meant to be implemented at some point? Should I use a workaround instead of get_features_name_out?

code snippets:

from skrub.datasets import fetch_employee_salaries
from sklearn.compose import make_column_transformer
from skrub import GapEncoder

datasets = fetch_employee_salaries()
X, y = datasets.X[:500], datasets.y[:500]

preprocessing = make_column_transformer(
    (GapEncoder(n_components=100), "division"),
)

preprocessing.fit(X)
preprocessing.get_feature_names_out()

and for dropcols:

from skrub.datasets import fetch_employee_salaries
from sklearn.compose import make_column_transformer
from skrub import DropCols
from sklearn.linear_model import LinearRegression
from sklearn.pipeline import make_pipeline

datasets = fetch_employee_salaries()
X, y = datasets.X[:500], datasets.y[:500]

model = make_pipeline(DropCols(['gender', 'department', 'department_name', 'division',
       'assignment_category', 'employee_position_title', 'date_first_hired']), LinearRegression())

model.fit(X, y)
model.get_feature_names_out()

They both output an error.
Thank you :)!

[GaelVaroquaux]

I believe that they should be there and that this is a bug.

We might need to think about our inheritance diagram, or a good base class.

I would put this as a bug, but not a blocker for release.

[rcap107]

The problem with DropCols is that get_feature_names_out() needs to take one additional argument that should be ignored

    def get_feature_names_out(self, input_features=None):
        check_is_fitted(self, "_columns")
        return self._columns

SelectCols and ApplyToCols all have the same issue

[rcap107]

The GapEncoder has a similar but inverse problem

get_feature_names_out takes the parameter n_labels, and when it's being fitted everything works fine

but when it's called by the ColumnTransformer then n_labels is replaced by the column names and that raises the exception

so get_feature_names_out should have a different signature, like so

    def get_feature_names_out(self,input_features=None, n_labels=3):

[Vincent-Maladiere]

Also, the logic of the DropCols.get_feature_names_out is incorrect. It returns self._columns, whereas this variable represents the list of columns to drop, not the list of columns to keep.

[rcap107]

Good catch

[rcap107]

For clarity: the current version of DropCols returns the dropped columns when calling get_feature_names_out. In the fixed version, it will return the surviving columns instead (the inverse set).