Feature request: OneHotEncoder copying Feature names

[d-kleine]

Hi,

I would like to ask if it would be possible to add a feature to sklearn's OneHotEncoder:

1. to automatically create feature names in such way columns were named in the input data
2. one-hot encoded features are replacing input columns in original dataframe
3. that values of one-hot encoded features should always be 0 and 1 (easier to read; currently, they are floats 0.0 and 1.0)

Describe the workflow you want to enable

I would like to have an more comfortable way to applying one-hot enconding as implemented in pdpipe (using sklearn OHE):

#example data
import pandas as pd
df_example = pd.DataFrame({'groups': ["test1", "test2", "test3", "test4"],
                          'population': [10000, 20000, 30000, 40000]})
df_example

# copying example data for demonstation
df_example_pdp = df_example.copy() 
df_example_skl = df_example.copy()

# pdpipe OHE 
import pdpipe as pdp 
onehot = pdp.OneHotEncode(['groups'], drop_first=False) 
df_example_pdp = onehot.fit_transform(df_example_pdp) 
df_example_pdp.head()

Describe your proposed solution

Output should look like this:

# sklearn OHE 
from sklearn.preprocessing import OneHotEncoder 
enc = OneHotEncoder(handle_unknown='ignore')
enc_fitted = enc.fit_transform(df_example_skl[['groups']])
column_name = enc.get_feature_names(['groups'])
enc_df = pd.DataFrame(enc_fitted.toarray().astype(int), columns=column_name)
df_example_skl = df_example_skl.join(enc_df) 
df_example_skl.head()

Describe alternatives you've considered, if relevant

pdpipe provides an excellent way: features will be one-hot encoded and columns will be replaced with one-hot encoded features. Great and simple labeling for understanding, integers are easier to read than floats, small and easy-to-read syntax. Hope to see this in sklearn's OneHotEncoder directly.

Additional context

Tested with sklearn's newest version 0.23.2 and older

[NicolasHug]

Hi,

to automatically create feature names in such way columns were named in the input data

Feature names propagation is tracked in scikit-learn/enhancement_proposals#48 and related PRs / issues

one-hot encoded features are replacing input columns in original dataframe

Unlikely to happen: transform doesn't change the data inplace. We're thinking about outputting dataframe-like objects (see other SLEPs), but that's slightly different

that values of one-hot encoded features should always be 0 and 1 (easier to read; currently, they are floats 0.0 and 1.0)

You can select a dtype with the dtype parameter. However, if you use the OHE in a ColumnTransformer and the final array has a mix of real-valued data and integer data, it will be upcasted to floats.

Note that for now, numpy array are first class citizens of scikit-learn, not dataframe.