The Problem of Redundant Variables in Random Forests

Random forests are currently one of the most preferable methods of supervised learning among practitioners. Their popularity is influenced by the possibility of applying this method without a time consuming pre processing step. Random forests can be used for mixed types of features, irrespectively of their distributions. The method is robust to outliers, and feature selection is built into the learning algorithm. However, a decrease of classification accuracy can be observed in the presence of redundant variables. In this paper, we discuss two approaches to the problem of redundant variables. We consider two strategies of searching for best feature subset as well as two formulas of aggregating the features in the clusters. In the empirical experiment, we generate collinear predictors and include them in the real datasets. Dimensionality reduction methods usually improve the accuracy of random forests, but none of them clearly outperforms the others.