Why use one machine learning algorithm when you could use all of them? Superlearning is a technique for prediction that involves combining many individual statistical algorithms (commonly called “data-adaptive” or “machine learning” algorithms) to create a new, single prediction algorithm that is expected to perform at least as well as any of the individual algorithms. The superlearner algorithm “decides” how to combine, or weight, the individual algorithms based upon how well each one minimizes a specified loss function, for example, the mean squared error (MSE). This is done using cross-validation to avoid overfitting.
The motivation for this type of “ensembling” is that a mix of multiple algorithms may be more optimal for a given data set than any single algorithm. For example, a tree based model averaged with a linear model (e.g. random forests and LASSO) could smooth some of the model’s edges to improve predictive performance. Superlearning is also called stacking, stacked generalizations, and weighted ensembling by different specializations within the realms of statistics and data science.
Read more at:
/Service Ventures Team
Commentaires