Christian Lorentzen

A question for my understanding: Isn't the code executing 2 things in parallel, the `n_jobs` of `cross_val_score()` and the implicit `n_threads` of `HistGradientBoostingRegressor`. So the question is, are the number...

I am not sure if your benchmarking is as informative as you wish for: - What are the absolute timings (not just relative)? - Parallelism in small `n_samples` is not...

> do you think you could drive this? Yes, I can, but that's not the main problem: It needs community support (from scikit-learn and narwhals) and we (scikit-learn) need to...

> binning + MAE criterion: >…I'm pretty sure histogram based split is not compatible with MAE criterion I don’t think this is an obstacle. The features X are binned, not...

@betatim Your summer student still has to learn about forests and trees. Random forests like very random deep trees, gradient grown forests like shallow trees. LightGBM's default parameters have gradient...

Given that we already have so many examples, see #27151, I‘d prefer to have this here integrated into an existing one.

Thanks for putting in an existing example. I am a bit undecided between the one you chose "Probability Calibration curves"/plot_calibration_curve.py and "Probability Calibration for 3-class classification"/plot_calibration_multiclass.py. What do other maintainers...

I see the offending G^-1 for tiny alpha. Can we instead compute alpha * G^-1 and compensate for this change later on? This way, we could keep the reduced SVD.

What calculation step is it exactly that struggles with the numerical stability for tiny alpha?

While I think the math is correct, our doc certainly is not: - [SGD Math](https://scikit-learn.org/stable/modules/sgd.html#mathematical-formulation) has L2 penalty with and without the factor 1/2. - [Online Ine Class SVM Math](https://scikit-learn.org/stable/modules/sgd.html#online-one-class-svm)...