Add custom imputation strategy to SimpleImputer #28053
Conversation
mark-thm
force-pushed
the
me/custom-imputation2
branch
from
2c960f9 to
64eb8bf
Compare
There was a problem hiding this comment.
Thanks for the quick turnaround. What I'd imagined was strategy=np.maximum in which case we apply self.strategy(masked_X_column.compressed()) for each column, or, in an optimisation, identify that self.strategy is a ufunc and apply the function vectorized across columns. I think it's okay, initially, to expect that the callable take a dense 1d array as input.
Co-authored-by: Joel Nothman <joeln@canva.com>
sklearn/impute/_base.py
Outdated
| else: | ||
| raise RuntimeError(f"Unknown strategy {strategy}") |
There was a problem hiding this comment.
I don't think this can ever happen since we validate parameters before getting here and people shouldn't be calling these private methods themselves.
There was a problem hiding this comment.
I was unable to get code coverage to pass without adding the final condition and this test.
There was a problem hiding this comment.
If you remove this line, codecov doesn't have them to want to cover in the first place.
sklearn/impute/_base.py
Outdated
| else: | ||
| raise RuntimeError(f"Unknown strategy {strategy}") |
Reference Issues/PRs
#27986
What does this implement/fix? Explain your changes.
Adds a 'custom' strategy to
SimpleImputerthat enables supplying ones own statistics to produce an imputation value.In my experience, it's useful to be able to compute, for instance, minimum and maximum values of the inputs in addition to, for instance, mean, and this enables unifying the imputation logic and producing a single location to manage all imputations.
Any other comments?
I proposed a similar change in #27986 and @adrinjalali and @jnothman requested to see a variation that accepts a callable instead of explicitly supporting new statistics.