modules on the basis of how they work rather than function

I don't like this (even though we've done that in the past - inconsistently. Why is LDA and LinearSVC not in linear models?)

I'm +.5 for sklearn.impute, possibly moving when when we add the next class (KNNImputer I guess?).

Actually, given that MICE is not that far away, should be +1

MICE is not far away at all. sklearn.impute would be useful, but importing it would conceivably import neighbors, linear_model, ensemble and tree if we had implementations of MICE, KNN and forest-based imputation there. We have decidedly scattered anomaly detection around the place. I am uncomfortable about putting MICE and KNNImputer in preprocessing, but I'm not entirely certain that sklearn.impute is the right solution.

If we make sklearn.impute, do we rename Imputer to sklearn.impute.BasicImputer or FeaturewiseImputer or some such?

Of course we could just make KNNImputer live under neighbors and MICE live under ?ensemble.

I've wondered whether in some ways it would make sense to have a pseudo-module sklearn.classifiers, sklearn.regressors, sklearn.imputers, etc, that import from the relevant implementation locations...

Yeah... I would prefer a semantic organization, but it's not how we have done things in the past. I guess you suggest having that in parallel to the current structure? I wouldn't be opposed, but it's a big change. Is the goal to keep two places to import from long-term? That seems slightly confusing....

True, we have done a really weird mix. The fact that we have SGDClassifier which implements many losses, and LogisticRegression which implements many solvers shows that we're not the best with the consistency ;)

The MissingnessIndicator (#8075) should also live in this module, which may help users.

I'd like another opinion, but I think this should happen. Make sklearn.impute and move Imputer and all the open imputation-related PRs to this module.

I've opened this to contributors. Please

• copy sklearn/preprocessing/imputation.py to sklearn/impute.py, as well as the corresponding tests,
• deprecate Imputer in sklearn/preprocessing/imputation.py to be removed in v0.22
• create a sklearn.impute section in doc/modules/classes.rst
• update the deprecated section at the bottom of doc/modules/classes.rst
• update sklearn/__init__.py's __all__
• move imputation documentation from doc/modules/preprocessing.rst to doc/modules/impute.rst
• we might also want to rename Imputer in the new module to SimpleImputer, DummyImputer or something (ideas??)
• update any references to Imputer (in sklearn/, doc/ or examples/) to refer to the new location

and after merge, please advise contributors at #8075, #8478, #9212 to move their work to the new module.

Re: naming, I like ConstantImputer because it fills in all missing values in a feature with a constant, but it might not be clear from the name that is what it's doing. Maybe BasicImputer or NaiveImputer.

I agree because DummyRegressor is not actually useful for doing work, but the Imputer is quite useful.

@jnothman I can work on this. Thanks for the detailed steps.

@krishnakalyan3 Hey! If you have not done significant progress to this PR, I would like to address this issue. I have been working on this as well and will be opening PR in a while.

You'd be welcome to find a way to split up the work. You do have a few things still on your plate elsewhere, @thechargedneutron

@thechargedneutron ah you beat me to it. Thanks for taking up the task!.

@krishnakalyan3 Sorry mate!! See you around 😄