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.

Just for the context, refer to scikit-learn/scikit-learn#23734 (comment)

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 :)

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

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

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 ".

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"?

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.

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.

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

Agreed with @amueller. Closing then.