Useful `scikit-learn` transformers for working with Pandas DataFrames

Description on Towards Data Science.

Quick --- What is Here?

Code for processing pandas.DataFrame that is compatible with sklearn.pipeline
High code coverage by unit tests (via pytest)
Continuous integration via GitHub Actions as I add more functionality

Motivation

Make scikit-learn pipelines retain and remember metadata, e.g. column names, from pandas DataFrames. Facilitates model debugging and interpretation! For details, see my article on Towards Data Science.

What about `ColumnTransformer` in scikit-learn?

To be fair, there is one way that scikit-learn utilizes metadata in DataFrames: ColumnTransformer can identify DataFrame columns by their string names, and directs your desired transformers to each column. Here is an example by Allison Honold on Towards Data Science.

Unfortunately, ColumnTransformer produces numpy arrays or scipy sparse matrices. My code extends ColumnTransformer such that it produces pandas.DataFrame as well.

Functionality Implemented

All files are located within directory src.

ImputeByGroup.py
- ImputeNumericalByGroup; groupby, calculate per-group statistics (e.g. median) of a numerical column with missing values, and then impute each group using said statistics.
- ImputeCategoricalByGroup; like the above, but for imputing discrete, categorical columns. Fills up missing values using the most frequent unique value.
PandasColumnTransformer.py
- Wrapper around sklearn.compose.ColumnTransformer for automatic bookkeeping of column names, even when the number of columns changed after a transformation (e.g. one-hot encoding)

Usage

Copy the source file(s) of interest from the src directory into your own project, and then import as necessary. Please see playground.ipynb for a usage demonstration.

openerror/PandasTransform