Pipelines using pandas transform output do not work with feature creation

[admivsn]

Describe the bug
Pipelines break when you use pandas transform output with feature creation.

To Reproduce
Steps to reproduce the behavior:

import pandas as pd

from feature_engine.creation import RelativeFeatures
from sklearn.pipeline import make_pipeline
from sklearn.dummy import DummyClassifier
from sklearn import set_config

set_config(transform_output="pandas")

X = pd.DataFrame({"feature_1": [1, 2, 3, 4, 5], "feature_2": [6, 7, 8, 9, 10]})
y = pd.Series([0, 1, 0, 1, 0])

pipeline = make_pipeline(
    RelativeFeatures(
        variables=["feature_1"],
        reference=["feature_2"],
        func=["div"]
    ),
    DummyClassifier()
)

pipeline.fit(X, y)
pipeline.predict(X)

ValueError: Length mismatch: Expected axis has 2 elements, new values have 3 elements

[solegalli]

How annoying!

Thank you for reporting @admivsn

I need to look into the inner workings of the pipeline.

By the looks of the error, the pipe is surprised to get the extra feature created by the RelativeFeatures.

I'll see if I can dig some time over the weekend.

Cheers

[ClaudioSalvatoreArcidiacono]

The root cause of this issue is that at this line here:

class RelativeFeatures(BaseCreation):
    ...
    def transform(self, X: pd.DataFrame) -> pd.DataFrame:
        X = super().transform(X)

The transform method of BaseCreation is called. This method will then call the get_features_out method defined in GetFeatureNamesOutMixin which will return the feature names out of the child class RelativeFeatures instead of the parent class BaseCreation and since BaseCreation.transform(self, X) returns a DataFrame with 2 columns and get_features_out returns 3 columns this mismatch causes the error described in the issue.

[solegalli]

Yes, I just figured that out. I am fixing as we speak. PR coming in the next minutes. Thank you!

That was a nasty one!