SUPERLEARNER Procedure
Super Learner Algorithm
Given an input data set with a vector of predictors 
and a response variable y, the super learner model consists of two layers of models that are used to describe the relationship between and y.
The first layer includes a library of B individual base learner models:
where 
are the model parameters for base learners 1 through B.
The second layer contains the meta-learner model m, which is a function of the individual base learners in the first layer:
where 
is the vector of meta-learner model parameters, also known as the super learner model coefficients.
Training a super learner amounts to estimating the model parameters in the base learner models and the meta-learner model. You estimate the base learner model parameters by fitting each base learner to the entire input data set. You estimate the super learner model coefficients by following these general steps:
Randomly split all observations in the input data set into K disjoint blocks.
Train each base learner by using only data in blocks 2 through K, excluding block 1.
For each trained base learner, generate predictions for block 1.
Repeat steps 2 and 3 (
) times, each time leaving out block 
for predictions. This step is known as k-fold cross-validation. Across all blocks, fit a regression of the observed responses on the out-of-sample predictions from step 3. This is known as the meta-learning step.
For more information about k-fold cross-validation, see the section K-Fold Cross-Validation. For more information about estimating the super learner model coefficients , see the section Meta-learning Methods.