ENH: sklearn.model_selection.ShuffleSplit compatible wit pandas.MultiIndex

[jeiglsperger]

Feature Type

• Adding new functionality to pandas

• Changing existing functionality in pandas

• Removing existing functionality in pandas

Problem Description

I wish I could smoothly apply sklearn.model_selection.ShuffleSplit with a pandas.Dataframe that has a pandas.MultiIndex. I tried to apply it, but it throws some weird KeyError.

Feature Description

This is my pseudo code how I would expect it to work:

from sklearn.model_selection import ShuffleSplit
from sktime.datatypes import get_examples
df = get_examples(mtype="pd-multiindex", as_scitype="Panel")[0]
splitter = ShuffleSplit(n_splits=3, random_state=42)
split = splitter.split(df.index.levels[0]) 
train_indexes = []
test_indexes = []
for train_index, test_index in split:
    train_indexes.append(train_index)
    test_indexes.append(test_index)
x_train, x_test = (df.loc[train_indexes[0]], df.loc[test_indexes[0]])

Alternative Solutions

Currently I only to a train_test_split as described here without any cross validation.

Additional Context

It would also be nice to extend this to the other split methods of sklearn.

[rhshadrach]

Thanks for the request.

I tried to apply it, but it throws some weird KeyError.

Can you provide a fully reproducible example? In what you've provided, df and x_train_val are missing.

[jeiglsperger]

Sorry for not providing a MWE. I edited my post accordingly. In the MWE I used example data from sktime, for which the code works. There the error (certain indexes are not being found) is somehow due to my dataset. Either way, I now use the workaround to handle ndarray in the preprocessing and then convert them to pd.DataFrame