FIX export_text and export_graphviz accepts feature and class names as array-like - #26289

@adrinjalali Tests added, but I'm not sure if the way I add them is neat enough.

we could call check_array on the feature names and actually accept all array-like data.

we could call check_array on the feature names and actually accept all array-like data.

Calling check_array ends up calling asarray on list of strings, which makes a copy:

from sklearn.utils.validation import check_array

a = ["first", "second"]
a_array = check_array(a, dtype=object, ensure_2d=False)
a_array[0] = "new_first"

print(a_array[0])
# new_first

print(a[0])
# first

I think list and ndarray covers most of the use case for export_text.

which makes a copy

In case of something like export_text, do we care?

In case of something like export_text, do we care?

There are a few ways to interpret this. If you mean:

1. "There should not be too many feature names" when using export_text, then I think it is a valid argument for making the copy.
2. "export_text is not performance critical", then I still care. I prefer not to use more resources if we do not have to.

Logistically, I think most users will pass in ndarray, list, or df.columns. Supporting df.columns leans toward converting things to an array. I am now +0.5 on converting the input to an ndarray.

@thomasjpfan @adrinjalali So now we are going for converting input to ndarray right? I also agree with that: this is consistent with what export_graphviz supports for now, and I will modify its docstring as well. Is there a suggestion on which function I should use to convert input to ndarray? I assume that there is something in sklearn.utils.validation but I did not find the specific one to use.

@Charlie-XIAO you can use utils.validation.check_array

@Charlie-XIAO you can use utils.validation.check_array

I remembered that @thomasjpfan said this would make a copy of lists, which is not desirable?

said this would make a copy of lists, which is not desirable?

From #26289 (comment), I think it's okay and make the copy.

@thomasjpfan How would the docstring be then? We cannot write "array-like" because array-like objects do not necessarily support __len__ right?

We cannot write "array-like" because array-like objects do not necessarily support len right?

If we convert inputs to ndarrays before using on them, then we can use "array-like".

@thomasjpfan Changes made, please let me know if tests need to be improved or the way I use check_array need to be changed. I will updated parameter validation of export_graphviz in #26034 after this change is merged. By the way, should this be FIX or ENH or MAINT?

@adrinjalali @glemaitre Thanks for your review! I've decoupled the tests, please let me know if it is okay (it has some repeated code). Otherwise I can use your other suggestion (i.e., just run twice since they are fast).

Oh right I completely forgot about parametrization lol, will do soon.

Thanks @Charlie-XIAO LGTM.