[python-package] feature_names_in_ spaces replaced by underlines incompatible with sklearn

[ThibaultDECO]

Fixes issue #7338

FIX feature_names_in_ whitespace replaced by underlines is incompatible with sklearn

[jameslamb]

Thanks for your interest in LightGBM and for taking the time to open a PR.

I've started CI for you without reviewing the changes closely yet, let's see what that says.

At a glance, I'll say I'm not very excited about having to track the original feature names everywhere. This is new complexity and that means new opportunity for bugs and maintenance burden 😫

I will review more closely once we've release v4.7.0 (#7129), that is my primary focus right now.

[jameslamb]

Because you're a first-time contributor, every CI run requires one of us to click a button. You may find it faster to run the tests locally.

cmake -B build -S .
cmake --build build --target _lightgbm -4
sh build-python.sh install --precompile
pytest tests/python_package_test/

[ThibaultDECO]

@jameslamb Thank you James, I will run the tests locally.
With regards to your pushback, I understand that it adds complexity but I don't see any other way around at this time.
Thank you for your review!

[jameslamb]

I don't see any other way around at this time.

Other options might be:

• raise an informative error closer to the boundary of the scikit-learn interface if any feature names have spaces (i.e. simply don't allow them)
• remove the normalization of feature names completely in Dataset construction (i.e. allow spaces)
• make the feature name normalization configurable via a new parameter, and have scikit-learn interface explicitly opt out of that when it constructs a Dataset
• stop supporting feature_names_in_ in the scikit-learn interface entirely (maybe it is just not worth it, we keep tripping over it here)

I think evaluating this will require coming up with a specific answer for the question "why does LightGBM remove spaces from feature names at all?", which will require looking through some project history.