I am implementing a custom estimator within the field of causality. Sparing you the details, each column in X is mapped to a "role" (e.g. treatment and control) and these "roles" define how the estimator works internally. This mapping of the columns of X to "roles" is passed in the estimator's init. Therefore, this estimator cannot work with arrays as there is no way to identify the role of each column. In this case, my estimator is failing some tests when using parametrize_with_checks.

What is the recommended solution? Potential options:

1. Your estimator cannot be sklearn compatible as it must accept arrays
2. Use some sort of tag?
3. Use the recently introduced expected_failed_checks in parametrize_with_checks?
4. Something else