[MRG] Initialize MissingIndicator with error_on_new = False #13974
Conversation
|
Please add a non-regression test. Thanks. |
|
I presume this also applies to IterativeImputer |
|
Would the test be to ensure that errors are thrown if relevant features are missing values? |
|
To test that no errors are raised if a feature has a missing value in test
but not in train, and that the transformed data is all finite.
|
…e IterativeImputer nonregression test and also default error_on_new to False.
|
@jnothman can you review? Thanks! |
|
Please revert the changes that are not related to the PR. |
|
Done! @thomasjpfan |
|
Please add an entry to the change log at |
|
Done! @jnothman. Thanks for the help/input. |
|
@thomasjpfan: change being pushed up. |
|
@NicolasHug I proposed putting this is an 0.21 series bug fix release, assuming we have a couple more fixes to join if (not sure if you want the recently raised divide by zero in gradient boosting to be fixed in 0.21.3 also, for instance) |
|
Agree this should be a bug-fix release. |
|
Implemented suggestions @NicolasHug @amueller |
|
Otherwise LGTM. |
7f50e82
into
scikit-learn:master
|
Thank you @fhoang7! |
Reference Issues/PRs
Resolves #13968
What does this implement/fix? Explain your changes.
Set error_on_new to False by default if add_indicator is True in order to suppress missing value errors in ignored features.
Any other comments?