Make feature selection transforms return columns in inverse_transform

[alex-hse-repository]

🚀 Feature Request

Make feature selection transforms return columns in inverse_transform

Motivation

Inverse transformation might be impossible to the other transforms because of the absence of some columns removed by feature selection

Proposal

• In BaseFeatureSelectionTransform and FilterFeaturesTransform

0. In constructor add flag return_features: bool=False
1. In method transform store the columns that will be remover in self._df_removed(only for return_features=True)
2. Add method inverse_transform, should return the columns form self._df_removed(only for return_features=True)

Test cases

For both transforms:

1. Method transform correctly store the data frame with removed columns
2. Method inverse_transform correctly return the removed columns
3. Backtest works correctly with pipeline including this transform(e.g transforms=[LogTransform("column_1"), FilterFeaturesTransform(exclude=["column_1"])])
4. We do not lose regressors if we do fit_transform -> inverse_transform -> transform
5. Flag return_features works correctly
6. Full test in Pipeline.backtest method

Alternatives

No response

Additional context

No response

Checklist

• I discussed this issue with ETNA Team