Default Value for SimpleImputer with all missing values

[vsocrates]

Describe the workflow you want to enable

In the Notes of the SimpleImputer, it says that:

Columns which only contained missing values at fit are discarded upon transform if strategy is not "constant".

It'd be nice to be able to set this to a default value in transform instead of dropping those columns if there is no valid summary statistic in fit.

I recognize that it may only be a couple of lines, but it seems useful to have the logic captured within the class instead of checking beforehand for all np.nans per column.

Describe your proposed solution

Near the lines

scikit-learn/sklearn/impute/_base.py

Lines 512 to 525 in 74bf394

	else:
	# same as np.isnan but also works for object dtypes
	invalid_mask = _get_mask(statistics, np.nan)
	valid_mask = np.logical_not(invalid_mask)
	valid_statistics = statistics[valid_mask]
	valid_statistics_indexes = np.flatnonzero(valid_mask)
	if invalid_mask.any():
	missing = np.arange(X.shape[1])[invalid_mask]
	if self.verbose != "deprecated" and self.verbose:
	warnings.warn(
	"Skipping features without observed values: %s" % missing
	)
	X = X[:, valid_statistics_indexes]

We could add something like

            invalid_mask = _get_mask(statistics, np.nan)
            valid_mask = np.logical_not(invalid_mask)
            valid_statistics = statistics[valid_mask]
            valid_statistics_indexes = np.flatnonzero(valid_mask)

            invalid_statistics_indexes = np.flatnonzero(invalid_mask) 
            ...
            X[:, invalid_statistics_indexes] = {default_value}
            # X = X[:, valid_statistics_indexes]

Describe alternatives you've considered, if relevant

An alternative is to do this logic before running the SimpleImputer.

Additional context

No response

[glemaitre]

It'd be nice to be able to set this to a default value in transform instead of dropping those columns if there is no valid summary statistic in fit.

So instead of dropping the columns, you would like to have some columns with constant values? If this is the case, what is the reason in practice?

[vsocrates]

The use case I came across was when creating a ByGroupImputer that would eventually concatenate all the groups together. Each group is relatively small and likely to contain different columns with all NaNs, so the final sizes of the imputed arrays is different. This would be solved by being able to fill in NaNs with 0 (or any filled value) instead of dropping them.

In general, having the input and output array sizes be the same after transform seems useful though.

[glemaitre]

But in terms of training a predictive model, you cannot do anything with these constant features?