Background: Random Forests is a popular classification and regression method that has proven powerful for various prediction problems in biological studies. However, its performance often deteriorates when the number of features increases. To address this limitation, feature elimination Random Forests was proposed that only uses features with the largest variable importance scores. Yet the performance of this method is not satisfying, possibly due to its rigid feature selection, and increased correlations between trees of forest.

Methods: We propose variable importance-weighted Random Forests, which instead of sampling features with equal probability at each node to build up trees, samples features according to their variable importance scores, and then select the best split from the randomly selected features.

Results: We evaluate the performance of our method through comprehensive simulation and real data analyses, for both regression and classification. Compared to the standard Random Forests and the feature elimination Random Forests methods, our proposed method has improved performance in most cases.

Conclusions: By incorporating the variable importance scores into the random feature selection step, our method can better utilize more informative features without completely ignoring less informative ones, hence has improved prediction accuracy in the presence of weak signals and large noises. We have implemented an R package “viRandomForests” based on the original R package “randomForest” and it can be freely downloaded from http://zhaocenter.org/software.