Deciding on an ensemble's composition

We described stacking as an advanced form of voting. Similarly to voting (and most ensemble learning techniques for that matter), stacking is dependent on the diversity of its base learners. If the base learners exhibit the same characteristics and performance throughout the problem's domain, it will be difficult for the meta-learner to dramatically improve their collective performance. Furthermore, a complex meta-learner will be needed. If the base learners are diverse and exhibit different performance characteristics in different domains of the problem, even a simple meta-learner will be able to greatly improve their collective performance.

Selecting base learners

It is generally...