✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: b06eb7c. Link to the linter CI: here}

It looks to me like all branches of the new code get exercised, I can’t tell why the coverage tool seems to think some elif statements don’t get hit but the blocks inside are fully covered.

I wonder if we should instead accept a callable which takes a column and returns a constant to fill the values.

I wonder if we should instead accept a callable which takes a column and returns a constant to fill the values.

That's probably a bit bigger of a change and adapts the interface to SimpleImputer quite a bit more -- notably it might be a bit awkward becuase:

• lambda authors would need to write both sparse and dense implementations
• the strategies today are all different kinds of stats (mean, median, most_frequent), it seems awkward to suddenly expand this list with 'lambda' or 'custom'

Besides the point, I'm happy to field questions about the min/max implementation and usage from a support perspective, but I'd definitely hesitate to support pairs of arbitrary lambdas.

What are the use cases for this?

For our models, the imputed value for some missing features performs best when set to minimum or maximum of the feature. Consider a feature which typically has a value of 0, provides highly negative signal when it's any positive integer, and for some of our inference-time executions is missing: in this case, we want to drive this feature with a value of 0 (so that our model doesn't weigh this feature heavily). We have a number of features where the right missingness value isn't something as obvious as 0, but we know it's "the typical minimum" or "the typical maximum", and rather than pre-compute this value by feature and store the constant, we'd rather 'train' the imputer to impute these features along with all the others.

What steps can I take to get this merged?

Thanks for the info, I'm happy for it to be included. But don't have bandwidth to review, no blocker from my side though.

So next would be for two reviewers to have a look. Unfortunately that's a bottleneck and we don't necessarily have enough reviewers for a timely review of all PRs. So we shall wait and see.

How hard do you think it would be to mock up a PR that takes a callable run over each array of non-missing values? Given how rarely the set of options here has had to change, I suspect that a generic solution will be more helpful for more users than this specific one...?

#28053