I am wondering why stacking uses strong learners as base learners.
How to understand it from expectation and variance way, or bias and variance way?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Weak learners are cheap to train, but have poor performance, however if you stack big number of them, they perform not worse, or better, then strong learners alone. Strong learners are usually expensive to learn and apply (slow, higher computational & memory requirements etc.), however stacking them would usually improve the performance above that each of the strong learners achieves alone. Stacking strong learners is probably one of the most popular strategies used on Kaggle competitions. On another hand, people rarely use it on production because the cost is high and usually the gain in performance is not that big comparably to either stacking weak learners, or using single strong learner.
