B.4.7 super

This learner combines the results of other learners into a single prediction. This two-step technique starts by applying a set of learners, called the sublearners, to part of the training dataset. The sublearners then make predictions on remainder of the training dataset, and these predictions are stored in a new, temporary dataset. The outputs for the temporary dataset are the same as the outputs for corresponding rows of the original training dataset. In the second step, a learner called the superlearner is trained on the temporary dataset. The superlearner is then applied to the testing dataset or to the held-out data points in a k-fold cross-validation experiment. This idea has been thoroughly investigated under the name stacking [45].

Keyword	Arg Type	Arg Vals	Default
Common
learners	string
superlearner	string		lr

Common keywords and arguments:

• learners string: This keyword allows specification of the sublearners. The string argument allows passing of sublearner names and arguments. The sublearner name is separated from its arguments by spaces, while the learners are separated from one-another by colons. This information must be parsed as a a single string, and your command shell may require quoting of the string argument. For example to specify Bayes Classifier and K-NN with $k=9$ as sublearners, one might use ``bc:newknn k 9'' as the string argument for the learners keyword. It is also possible to use learners saved with the train action. If the trained learner is saved in a file learnerfile, then it may be specified with the format ``FILE learnerfile'' as a sublearner.

• superlearner string: This keyword allows specification of the superlearner. The string argument allows passing of the superlearner name and arguments. The superlearner name is separate from its arguments by spaces. This information must be parsed as a a single string, and your command shell may require quoting of the string argument. For example to specify K-NN with $k=9$ as the superlearner, one might use ``newknn k 9'' as the string argument. Note that it is not possible to load the superlearner from a file saved by the train action using the superlearner keyword; please use the predict action instead.