In principle sample_weight can be accessed when fitting TargetEncoder. So at first sight this is something that can be considered as a new feature. There seems to be some support for it as well (👍 on the issue).

Related to #11316 for testing.

I am working on this.

Hi! I am working on this issue with @MiguelParece. We have already incorporated sample weights to compute the weighted averages, but we are currently uncertain about whether we should also use them to calculate the variance when smooth="auto". I believe using sample weights for variance calculation is the standard approach, but we would like to confirm if this is indeed the best practice before proceeding with the implementation and testing.

@mayer79 @betatim

Hi! I am working on this issue with @MiguelParece. We have already incorporated sample weights to compute the weighted averages, but we are currently uncertain about whether we should also use them to calculate the variance when smooth="auto". I believe using sample weights for variance calculation is the standard approach, but we would like to confirm if this is indeed the best practice before proceeding with the implementation and testing.

@mayer79 @betatim

Good point - I'd go for a weighted variance. Now, there are different ways to calculate weighted variances, see e.g. https://numpy.org/doc/stable/reference/generated/numpy.cov.html, which is related to the meaning of the "weights":

• frequency weights: a value 2 means "the row represents two observations" or
• general weights, where a value 2.5 could mean: "the observation is very important"

What type of weights should we stick to @lorentzenchr ?

Hi! I am working on this issue with @MiguelParece. We have already incorporated sample weights to compute the weighted averages, but we are currently uncertain about whether we should also use them to calculate the variance when smooth="auto". I believe using sample weights for variance calculation is the standard approach, but we would like to confirm if this is indeed the best practice before proceeding with the implementation and testing.
@mayer79 @betatim

Good point - I'd go for a weighted variance. Now, there are different ways to calculate weighted variances, see e.g. https://numpy.org/doc/stable/reference/generated/numpy.cov.html, which is related to the meaning of the "weights":

* frequency weights: a value 2 means "the row represents two observations" or

* general weights, where a value 2.5 could mean: "the observation is very important"

What type of weights should we stick to @lorentzenchr ?

From what I understand from this comment on another issue regarding sample_weight we should use frequency weights. Would you agree @mayer79 ?

In the end, it should not matter that much. I prefer the general weights as those are met more often in ML and freq weights are just more special (like in a survey).

Agree @lorentzenchr . Maybe mention in the docstring of the new argument which formula is used to calculate weighted variance.

check_sample_weight_equivalence_on_dense_data which is part of our common test suite checks for the frequency weights semantics. I think this is what we should strive for in general but if there are cases where those semantics are not reachable for a particular reason, this reason should be made explicit in the XFAIL message of that particular estimator:

Related PR / discussion: #30564