A Study of Strength and Correlation in Random Forests

Simon Bernard (1) , Laurent Heutte (1) , Sébastien Adam (1)

1 DocApp - LITIS - Equipe Apprentissage

In this paper we present a study on the Random Forest (RF) family of classification methods, and more particularly on two important properties called strength and correlation. These two properties have been introduced by Breiman in the calculation of an upper bound of the generalization error. We thus propose to experimentally study the actual relation between these properties and the error rate in order to confirm and extend the Breiman theoretical results. We show that the error rate statistically decreases with the joint maximization of the strength and minimization of the correlation, and this for different sizes of RF.

Mots clés: Random Forests; Ensemble of Classifiers; Strength; Correlation

Domaines: Apprentissage [cs.LG]

Simon Bernard, Laurent Heutte, Sébastien Adam. A Study of Strength and Correlation in Random Forests. International Conference on Intelligent Computing, Aug 2010, Changsha, China. pp.186-191, ⟨10.1007/978-3-642-14831-6_25⟩. ⟨hal-00598466⟩