skrub-data.org/stable/refer...

skrub-data.org/stable/refer...

Example: skrub-data.org/stable/auto_...

For even more control over column selection, skrub provides a collection of selectors that let you partition dataframes by data type, column name, or user-specified functions.

All these transformers can be concatenated and inserted in a scikit-learn pipeline to build a feature matrix with complex column selection operation, and can be seen as an alternative for the scikit-learn ColumnTransformer.

ApplyToFrame selects columns in the same way, but then uses all of them at the same time as input to the transformer: this is useful for dimensionality reduction.
SelectCols and DropCols can be used as "filtering blocks" in a pipeline.

Slides:
skrub-data.org/skrub-materi...

Thanks to @riccardocappuzzo.com , @glemaitre58.bsky.social and Jérôme Dockès for preparing the talk, and mentoring at the sprint!

The sprint was also a big hit, with both new and old contributors working on issues and getting to know the repository.

And to cap it all off, thanks to P16 we have stickers now 🚀

What a banger is skrub @skrub-data.bsky.social !

Big thumbs up for the sklearn team & the maintainer of this package

🛠️ Main bugfixes
- Fixed the display of DataOp objects in Google Colab cell outputs.
- Fixed the range from which choose_float and choose_int sample values when log=False and n_steps is None.
- The SkrubLearner used to do a prediction on the train set during fit(), this has been fixed.

👀 Changes and deprecations
- Ken embeddings are now deprecated.
- The accepted values for the parameter how of .skb.apply() have changed. The new values are "auto", "cols", "frame", and "no_wrap".
- The parameter splitter of .skb.train_test_split() has been renamed split_func.

🚀 New features
- The DataOp.skb.full_report() now displays the time each node took to evaluate.
- The User guide has been reworked and expanded.

Here's another example on how to tune ML models with skrub Data Ops: skrub-data.org/stable/auto_...

The plot in the video was created for our EurosciPy 2025 tutorial on forecasting time series: skrub-data.org/EuroSciPy202...

The plot is interactive: you can select a range of results, and it will highlight only the runs within that range, enabling you to refine your search further. It also tracks fit and score times, so you can identify which parameters most impact runtime.

skrub-data.org/stable/refer...