Rotation Forests for regression

• The performance of the Rotation Forest ensemble method for regression is analysed.
• The experimental validation uses ensembles of regression trees and 61 datasets.
• Rotation Forest has better results than Bagging, Random Subspaces and AdaBoost.R2.
• Diversity-error diagrams show the behaviour of the ensemble methods.

Rotation Forest, originally proposed for the combination of classifiers, has shown itself to be very competitive, when compared with other ensemble construction methods. In this paper, the performance of Rotation Forest for combining regressors is investigated using a broad range of datasets, 61 in total, which vary in size from 13 to more than 40,000 instances, and from 2 to 60 attributes, with both numeric and nominal attributes. Rotation Forest has favourable results when compared with Bagging, Random Subspaces, Iterated Bagging and AdaBoost.R2, according to average ranks and a scoring matrix. Diversity error diagrams are used to analyse the behaviour of the ensemble methods.