Margin based variable importance for random forest

Random forest variable importance measure (RF-VIM) has been widely used in many applications such as bioinformatics. The permutation based VIM accesses the variable importance with changes of the OOB accuracy caused by random permutation. In this paper a margin based variable importance measure was proposed, which use the changes in the margin distribution of training samples caused by random permutation to assess the importance of features. Experiments on gene expression data show that the method outperforms both the original RF-VIM and the Gini importance on some datasets.