In imbalanced-learn we have cases (e.g. scikit-learn-contrib/imbalanced-learn#157) where some samplers like RandomUnderSampler and RandomOverSampler could naturally accept NaNs for X because the values of the X are irrelevant to the random sampling itself. If we modifiy the Random*Samplers to accept NaNs the check_estimator tests fall for them.

So we have two options:

• Skip the check_estimator tests for the Random*Samplers
• Do not accept the NaNs at all and document it clearly.

Any workaround on this? What are your thoughts?

This issue in general is related to #6981.