scikit-learn: Tests are failing with the new version of lightgbm (3.0.0)

The Linux pylatest_pip_openblas_pandas job is failing with the new version of lightgbm (3.0.0).

I am opening this issue for investigating the underlying problem and see what is wrong.

About this issue

Original URL
State: closed
Created 4 years ago
Reactions: 1
Comments: 21 (21 by maintainers)

Most upvoted comments

@glemaitre I don’t think so as the encoding in ESL and LogitBoost (J classes) in “Additive Logistic Regression” use indicator targets y_j in {0, 1} and y_j=1 iff target is class j.

If my analysis above is correct, I’d propose

to adapt the learning rate when comparing scikit-learn with LightGBM in the multiclass case
to open a separate PR to investigate if the symmetrization step is beneficial.

lorentzenchr on Sep 4, 2021

@lorentzenchr speaking with @ogrisel, we thought to ping you on this one 😃 With your PR to share the loss functions in the future, it means that the categorical cross-entropy would impact as well the LogisticRegression at least in the multinomial formulation. Did you ever encounter this normalization beforehand?

glemaitre on Sep 2, 2021

Maybe we could just pin lightgbm ❤️.0 in CI to fix master for now, and then investigate without rush?

rth on Sep 1, 2020