Does BorutaPy work with cuML RandomForestClassifier? about boruta_py HOT 8 OPEN

At the moment boruta tries to set the random state to all estimators. cuML's RF classifier do not have this parameter.

You can try a fix like with lightgbm. Something like this before the else part could help you.

if isinstance(self.estimator, cuml_type_here): pass

from boruta_py.

Thanks @Wuuzzaa.

I made the adjustment you recommended but now I'm receiving this error: "ValueError: Only methods with feature_importance_ attribute are currently supported in BorutaPy."

Any recommendations on this issue?

from boruta_py.

Seems like the implementation from cuML´s random forest differs quiete a lot from sklearns. I just took a look at the docu and do not found something similar to the feature importance.
cuML Random Forest

Some kind of feature importance is necessary for boruta to determine which features are useful. I think there is no easy way to work around this issue.

from boruta_py.

@curtisraymond and @Wuuzzaa Hi ... any solution for this?

I'm going through the same problem. However, I'm getting a different error: "integer required"

Error

TypeError Traceback (most recent call last)
/usr/local/lib/python3.10/dist-packages/boruta/boruta_py.py in _get_imp(self, X, y)
383 try:
--> 384
385 self.estimator.fit(X, y)
randomforestclassifier.pyx in cuml.ensemble.randomforestclassifier.RandomForestClassifier.fit()

TypeError: an integer is required

ValueError: Please check your X and y variable. The providedestimator cannot be fitted to your data.
an integer is required

from boruta_py.

My blind guess would be an error on your y data? y must be integers. Did you check your X and y for compatible Data types.
For the types see: docu

from boruta_py.

Hi @Wuuzzaa ..

Thank you for the quick reply.

y are integers. It works fine when I use sklearn's RF classifier. But I get this error when I use cuML's RF classifier.

My guess is that there might be an incompatibility between cuML and BorutaPy

from boruta_py.

BorutaPy was never planned to be used within cuML. Seems like it still do not work. Like beckernick mentioned there is still an open Issue on cuML for the implementation of the Feature Importance which is needed for boruta to work.

from boruta_py.

Thanks for linking that issue @Wuuzzaa !

@lindeberg25 , we'd love to learn more about your use case and performance impact of using cuML's Random Forest vs. scikit-learn's RF. Let's continue the discussion on the linked issue.

from boruta_py.