sklearn-cv-pandas

RandomizedSearchCV/GridSearchCV with pandas.DataFrame interface

Why do I want this?

• I usually prepare features as pandas.DataFrame
• Scikit learn input should be array-like. https://scikit-learn.org/stable/glossary.html#term-array-like.
• Although it includes pandas.DataFrame, there are some issues;
  • It does not support Int64 data type
  • Output model does not remember which columns should be used

Solution

• Provide GridSearchCV / RandomizedSearchCV with pandas.DataFrame interface
  • Internally preprocess DataFrame to be applicable for sklearn
• Output of fit command is now original Model object, which
  • stores column name information
  • provides pandas.DataFrame interface for prediction

Installation

pip install sklearn_cv_pandas

Development

Requires Python 3.11+ and uv.

uv sync --group dev
uv run pytest

Usage

Configure CV object

Instantiate CV in the same manner as original ones.

from scipy import stats
from sklearn import linear_model
from sklearn_cv_pandas import RandomizedSearchCV

estimator = linear_model.Lasso()
param_dist = dict(alpha=stats.loguniform(1e-5, 10))
cv = RandomizedSearchCV(estimator, param_dist, scoring="neg_mean_absolute_error")

fit with pandas.DataFrame

Our CV object has new methods fit_holdout_pandas and fit_cv_pandas. Original ones requires x and y as numpy.array. Instead of numpy array, you can specify one pandas.DataFrame and column names for x (feature_columns), and column name of y (target_column).

model = cv.fit_cv_pandas(
    df, target_column="y", feature_columns=["x{}".format(i) for i in range(100)], n_fold=5
)

predict with pandas.DataFrame

You can run prediction with pandas.DataFrame interface as well. Output of fit_holdout_pandas and fit_cv_pandas stores feature_columns and target_column. You can just input pandas.DataFrame for prediction into the method predict.

model.predict(df)