[python-package] Support feature_names_in_ attribute via sklearn API

[ravwojdyla]

Summary

sklearn API supports feature_names_in_ attribute on a fitted model (SLEP007), which remembers the feature names/columns that went into the model.fit method. This can be very useful information, and is a standard worth conforming to. Afaiu right now that information is available in the booster:

est.booster_.feature_name()

It shouldn't be too hard to conform to also expose that information via feature_names_in_ attribute 🙏

Motivation

It would conform to the sklearn API standards, improve usability of LightGBM models, especially when used along with other sklearn models and Pipelines.

References

• https://scikit-learn-enhancement-proposals.readthedocs.io/en/latest/slep007/proposal.html

[jameslamb]

Thanks for using LightGBM and taking the time to report this!

We'd welcome this addition, would you like to contribute it?

---

And a side question.... do you think it's an oversight that scikit-learn's estimator checks don't enforce this? We follow https://scikit-learn.org/stable/modules/generated/sklearn.utils.estimator_checks.check_estimator.html in LightGBM's tests to try to catch such things

LightGBM/tests/python_package_test/test_sklearn.py

Lines 1285 to 1288 in 255c93b

	@parametrize_with_checks([lgb.LGBMClassifier(), lgb.LGBMRegressor()])
	def test_sklearn_integration(estimator, check):
	estimator.set_params(min_child_samples=1, min_data_in_bin=1)
	check(estimator)

Using scikit-learn==1.3.2 (the latest released version as of this writing), check_estimator() says LGBMClassifier and LGBMRegressor are compliant with scikit-learn's expectations for estimators.

import lightgbm as lgb
from sklearn.utils.estimator_checks import check_estimator

check_estimator(lgb.LGBMClassifier())
check_estimator(lgb.LGBMRegressor())

But in the SLEP you linked, it says the following:

Backward Compatibility
All estimators should implement the feature_names_in_ and get_feature_names_out() API. This is checked in check_estimator...

[nicklamiller]

I would very much like to contribute to LightGBM and this seems like a great issue, with @ravwojdyla's blessing, I'd be happy to make this contribution.

[ravwojdyla]

@nicklamiller sounds great - thank you!

[jameslamb]

Do either of you know the answer to my question about check_estimator() from the latest scikit-learn not complaining about this?

[nicklamiller]

Backward Compatibility
All estimators should implement the feature_names_in_ and get_feature_names_out() API. This is checked in check_estimator...

@jameslamb I agree that based on SLEP007, this functionality should be implemented in check_estimator and does not appear to be. Here's a somewhat recent issue of sklearn estimators that lack(ed) this attribute, it looks like the attribute it is only checked/created if missing when _validate_data is called.

I can open an issue in sklearn and propose this behavior is more rigorously checked with check_estimator.

[jameslamb]

Thanks very much for the link to scikit-learn/scikit-learn#27907 @nicklamiller !

Please link to this issue from whatever one you create in scikit-learn.