I am not really keen to drop missing values with an option. I would prefer some code allowing to make it more explicit than just an option.

We currently can do that:

import sklearn
from sklearn.preprocessing import FunctionTransformer, OneHotEncoder
from sklearn.pipeline import make_pipeline

sklearn.set_config(transform_output="pandas")


def _drop_None_cols(df):
    col_names = [col for col in df.columns if "None" in col]
    if len(col_names):
        return df.drop(columns=col_names)
    return df


encoder_dropping_None = make_pipeline(
    OneHotEncoder(sparse_output=False),
    FunctionTransformer(_drop_None_cols),
)
encoder_dropping_None.fit_transform(test_df)

I think it would make sense to have:

• handle_missing="encode_as_category" by default with an option to set it to handle_ignore="ignore",
• missing_marker=(np.nan, pd.NA, None) which make it possible to specify which values should be consider missing value markers.

The missing_marker parameter would only be used if handle_ignore="ignore" to avoid creating output columns for such values.

@glemaitre: Thanks for your answer. What you suggest is what I am actually doing already.
My question is about including this in the OneHotEncoder itself and thus avoiding the additional FunctionTransformer.

In order to avoid collinearity, I could use the parameter drop='first' but it seems more intuitive to drop the "_None" feature instead. Don't you agree? It is what the pandas.get_dummies() function does by default.

@ogrisel: Sorry, I'm not sure to understand what you are suggesting.

@woodly0 @ogrisel is suggesting that there should be a parameter to handle (null or None or pd. NA) values or we can give values that can be treated like None.

I am willing to work on that and based on @ogrisel comment I'm thinking of adding two parameters:

handle_missing : {'encode_as_category', 'ignore'}, default='encode_as_category'
    Specifies the strategy to handle missing values during encoding.

    - If "encode_as_category", missing values are treated as a separate category and encoded into distinct
      columns, regardless of the `missing_values` parameter.
    - If "ignore", missing values are excluded from the encoding process, and no column is created for them
      in the output.

missing_values : int, float, str, np.nan or None, default=np.nan
    The placeholder to identify missing values when `handle_missing="ignore"`. All values equal to
    `missing_values` will be considered missing and ignored during encoding.

    This parameter has no effect when `handle_missing="encode_as_category"`.

I chose the name missing_values for consistency with sklearn.impute, and the description aligns with their style.

I'll be waiting for feedback on that and will start working on it if it looks good to everyone.