sklearn Pipe Step Interface for vtreat

We’ve been experimenting with this for a while, and the next R vtreat package will have a back-port of the Python vtreat package sklearn pipe step interface (in addition to the standard R interface).

This means the user can express easily express modeling intent by choosing between coder$fit_transform(train_data), coder$fit(train_data_cal)$transform(train_data_model), and coder$fit(application_data).

We have also regenerated the current task-oriented vtreat documentation to demonstrate the new nested bias warning feature:

• Regression: R regression example, Python regression example.
• Classification: R classification example, Python classification example.
• Unsupervised data preparation: R unsupervised example, Python unsupervised example.
• Multinomial classification: R multinomial classification example, Python multinomial classification example.

And we now have new versions of these documents showing the sklearn $fit_transform() style notation in R.

• Regression: R $fit_transform() regression example.
• Classification: R $fit_transform() classification example.
• Unsupervised data preparation: R $fit_transform() unsupervised example.
• Multinomial classification: R $fit_transform() multinomial classification example.

The original R interface is going to remain the standard interface for vtreat. It is more idiomatic R, and is taught in chapter 8 of Zumel, Mount; Practical Data Science with R, 2nd Edition, Manning 2019.

In contrast, the $fit_transform() notation will always just be an adaptor over the primary R interface. However, there is a lot to be learned from sklearn’s organization and ideas, so we felt we would use make their naming convention available as a way of showing appreciation and giving credit. Some more of my notes on the grace of the sklearn interface in being a good way to manage state and generative effects (see Brendan Fong, David I. Spivak; An Invitation to Applied Category Theory, Cambridge University Press, 2019) can be found here.