MNT consolidate target type tags

[adrinjalali]

This consolidates the following tags:

binary_only (default=False)
    whether estimator supports binary classification but lacks multi-class
    classification support.

multilabel (default=False)
    whether the estimator supports multilabel output

multioutput (default=False)
    whether a regressor supports multi-target outputs or a classifier supports
    multi-class multi-output.

multioutput_only (default=False)
    whether estimator supports only multi-output classification or regression.

into

target_type (default=["multiclass", "single-output"])
    the supported target types of the estimator, which can be a subset of:
        - 'binary': binary classification. All estimators supporting
          'multi-class' also support 'binary'; classifiers only;
        - 'multi-class': multiclass classification
        - 'multi-label': multilabel output
        - 'multi-output': multi-output regression or classification, i.e. `y`
          has shape `(n_samples, n_outputs)`
        - 'single-output': single-output regression or classification, i.e. `y`
          has shape `(n_samples,)`

Towards #28910

cc @glemaitre @thomasjpfan

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 8e35307. Link to the linter CI: here}

[glemaitre]

I think that we should have an instance to explain that you can multi-output and not support a single output.

[adrinjalali]

Added a note:

Note that an estimator can support multi-output but not single-output, or vice
versa. The same goes for binary and multi-class support.

[glemaitre]

Together with the other comment, we might benefit of an example to show what we mean.

[glemaitre]

I think that we should have an instance to explain that you can multi-output and not support a single output.

[thomasjpfan]

I see the logic behind it, but it's unfortunate we can use the names in type_of_target.

[thomasjpfan]

If the estimator is a regressor, then having a "multiclass" tag feels strange to me.

[adrinjalali]

An alternative here is to have a set of binary flags, instead of a set of flags, as in, introduce 5 flags which are all binary. It makes updating the flags in inheritance slightly easier. Right now once we implement the inheritance for __sklearn_flags__, the user should do something like

    flags["target_type"] = set(flags["target_type"]) | {"multi-output"}

I'm not sure which one would be better.

[adrinjalali]

Another alternative is to have two flags: classification-type and output-shape/type maybe?

[thomasjpfan]

An alternative here is to have a set of binary flags, instead of a set of flags, as in, introduce 5 flags which are all binary.

I like the binary flags approach.

[adrinjalali]

If we go back to binary flags, then the only change I see necessary is to:

multioutput_only (default=False)
    whether estimator supports only multi-output classification or regression.

to

single_output (default=True)
    whether estimator supports single output.

Note that an estimator can support `single_output` and/or `multi_output`.

[thomasjpfan]

I'm okay with single_output. What do you think of having two binary tags for single-output classification? i.e. binary and multi-class.

• binary=True, multi-class=False == binary_only
• binary=False, multi-class=True == impossible
• binary=False, multi-class=False == does not support single output classification
• binary=True, multi-class=True == most classifiers

Although adding another tag is more complexity, I do not like how binary_only implies that multi-class is not support.

[adrinjalali]

binary=False, multi-class=False == does not support single output classification

I think single and multi output aspect is independent from binary/multi class aspect. You could be a multi output only classifier that only supports binary outputs for instance.