The main paper was published in 2010 and was cited more than 400 times so would meet our inclusion criteria.

https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=Super+Learner+In+Prediction&btnG=

The algorithm is presented in section 2 of the paper:

It seems very close to our StackingClassifier/Regressor model where the second stage model is a Ridge classifier or regressor model with positivity constraints on the coefficients (and an extra constraint that they should sum to one):

https://scikit-learn.org/stable/modules/ensemble.html#stacked-generalization

The main difference with what we currently have in scikit-learn is the preconfigured list of base estimators used to populate the first stage.

Things that we could explore:

• improve the existing usage example of stacking in scikit-learn to provide a better list of base models (e.g. mixing MLPs, tree-based models and linear models with non-linear feature engineering, e.g. splines and/or kernel approximators) and illustrate this on a more realistic dataset with a mix of categorical and numerical features;
• consider providing a preconfigured list of base estimators/pipelines as part of the libraries itself;
• maybe create a superlearner package under scikit-learn-contrib that tries to mimic the default config and high-level features of the R package?

Our only example is:
https://scikit-learn.org/stable/auto_examples/ensemble/plot_stack_predictors.html

Totally agreed that it is not to be implemented in Scikit-Learn. The idea would be to include it in the documentation through an example, either in the existing stacking example or through a new example. With an explicit mention of the SuperLearner paper and/or package in the example? (so that it is findable when looking for "super learner python" on your favorite search engine), wdyt?

your other suggestions are really interesting too, maybe in a second step if needed? depending on the reach of use of the example base code (not sure how to measure it directly, but communication metrics can help)

Providing a list of good based linear pipelines might also be useful for #6329 (greedy ensemble) which is very related.

I also agree that mentionning "SuperLearner" either in the docstring of the stacking meta-estimators, or in the user guide or in the example or in all of the above might be helpful for googleability.