DOC Rename 'default' to 'native' for set_output SLEP

[thomasjpfan]

Based on the discussion in https://github.com/scikit-learn/scikit-learn/pull/23734#pullrequestreview-1135967664, reviewers concluded that "default" was not descriptive enough. The PR was updated to use "native" instead.

[adrinjalali]

So "native" means convert to numpy.ndarray or scipy.sparse? I personally find "default" more descriptive in this case. I'm not sure what's native about numpy/scipy from the perspective of a user.

[glemaitre]

Just for the context, refer to https://github.com/scikit-learn/scikit-learn/pull/23734#discussion_r985071192

Both "default" and "native" are not passepartout names. "Native" would imply that this is the native format given by the estimator that could optionally depend on the inputs. I would feel that "default" should mean a single type of output.

But still, both terms are really generic and are used to define something that is ill-defined (if not reading the estimators' documentation).

Bottom line, I don't have a strong opinion and I don't have a better proposal :)

[adrinjalali]

Would "unchanged" be a better name there? As in, we don't try to change the output from whatever's produced by the transformer.

[glemaitre]

It was a proposal from @thomasjpfan and I would also be fine with it.

[lorentzenchr]

As I already said

I'm happy with any other default value than "default", which I would consider bad design because it does not tell anything about the actual behaviour.

That means I favor a string that gives a hint at what it actually does. The meaning is - please correct me: A scikit-learn transformer outputs numpy.ndarray or scipy.sparse. This does not guarantee anything about 3rd party transformers.

How about "ndarray_or_sparse"? For sure, I‘m also ok with "unchanged ".

[thomasjpfan]

With the Array API PR, scikit-learn can also output cupy.array_api.Array, so the more descriptive name is "array-like_or_sparse".

I do not really like "unchanged" anymore. For example, it leads to this UX:

# By default, `est` returns an ndarray
est.transform(X_df) # this is an ndarray

est.set_output(transform="pandas")
est.transform(X_df) # this is a dataframe

est.set_output(transform="unchanged")
est.transform(X_df) # this is a ndarray??

I originally choose "default" to mean:

• "what the default transformer outputs"
• "what a not configured transformer outputs"

Maybe... "not-configured", "original", or "initial"?

[lorentzenchr]

Some options for the default value of set_output with the meaning "output what you get out of estimator.transform() without modifying anything". Please add your choice.

nr	default value	advantage	disadvantage
1	"default"	concise, no SLEP change needed	unspecific: default of what?
2	"native"	concise	unspecific: native to what?
3	"unchanged"		transform does change the input, e.g. a df as input may output a ndarray
4	"not-configured"
5	"original"
6	"initial"
7	"array-like_or_sparse"	very specific	long (maybe wrong for 3rd party estimators)
8	"estimator_default"	clear meaning	long
9	None	pythonic default value	not an option as it is used as sentinel value internally

[lorentzenchr]

After a good 🏃, I withdraw my concerns for "default". If this option has consensus among the core devs, let's go with it and document well what it means exactly.

[amueller]

so... close and leave as-is? I think in the end none of the words easily captures what's going on and folks will have to read the docs

[glemaitre]

Agreed with @amueller. Closing then.