META - Improvements for the Data Ops

Open rcap107 opened this issue 3 months ago • 1 comments

This issue is intended to track improvements and fixes for the Data Ops that should be implemented in the short term.

[ ] Rework joins/move joins to the .skb namespace
[ ] Add set_data(): if I have a Data Op with empty values, this lets me go back to interactive mode
[ ] Improve performance when adding a data op (at the moment, the operation has quadratic complexity)
[ ] Allow passing kwargs to fit, predict, cross_validate
[ ] Add a way to reweigh choose_from so that certain branches are not chosen more often than others
[ ] Add a "verbose" parameter to .skb.cross_validate to show the search space used to train the model, maybe with a message saying "this was trained on default parameters"
[ ] Add a way to track the predict results in skrub.cross_validate and .skb.cross_validate. At the moment, it's very hard to export all the results of the cross-validation fold, but this is useful info for evaluating the performance of the models.
[ ] Add truncated_after to Data Ops. At the moment, it's only available for learners.

Sep 22 '25 07:09 rcap107

I added the section on examples with external libraries

Oct 07 '25 09:10 rcap107