flennerhag
commented
4 years ago The meta-learner is fitted using cross-validation, so the performance of the meta-learner depends on what class you are using and how much data you have. The original SuperLearner using a linear regression as the meta-learner. An even simpler method is to use majority voting (linear regression with weights fixed to 1).
when you create the ensemble, you pass in a score function. score-m is the mean value of score over cv-folds.
I recently started to use your Superlearner with some classifiers in the first layer and then the meta-learner. When I just do the first layer without the meta-learner there are some classifiers which perform way better in comparison to the case of the inclusion of the meta-learner?
How does the Meta-Learner Layer determine the weights of the previous layer?
Further, when i call the results you report "score-m ....", but which score is meant by this? When I calculate the score for the actual scorer i used i get something different? Thanks