Generally, the meta-learner should be a relatively simple machine learning algorithm, in order to avoid overfitting. Furthermore, additional steps should be taken in order to regularize the meta-learner. For example, if a decision tree is used, then the tree's maximum depth should be limited. If a regression model is used, a regularized regression (such as elastic net or ridge regression) should be preferred. If there is a need for more complex models in order to increase the ensemble's predictive performance, a multi-level stack could be used, in which the number of models and each individual model's complexity reduces as the stack's level increases:
Another really important characteristic of the meta-learner should be the ability to handle correlated inputs and especially to not make any assumptions about the independence of features from one another, as naive Bayes classifiers do. The inputs to the meta-learner (metadata) will be highly correlated. This happens because all base learners are trained to predict the same target. Thus, their predictions will come from an approximation of the same function. Although the predicted values will vary, they will be close to each other.