With random Subspaces, estimators differentiate because of random subsets of the features. Again, such a solution is achievable by tuning the parameters of BaggingClassifier and BaggingRegressor, by setting max_features to a number less than 1.0, representing the percentage of features to be chosen randomly for each model of the ensemble.
Instead, in Random Patches, estimators are built on subsets of both samples and features.
Let's now examine in a table the different characteristics of pasting, bagging, random subspaces, and random patches as implemented using the BaggingClassifier and BaggingRegressor in scikit-learn:
Ensembling |
Purpose |
Hyperparameters |
Pasting |
A number of models are built ... |