Figure 1.

The methodology scheme. In order to avoid the selection bias, an external resampling scheme (stratified partitioning) is coupled to an internal K-fold cross-validation applied to the ranking method under study. The modeling procedure is replicated on resampled versions of the original data set, with validation always operated on a test set disjoint from the development material (dev) for the current run (VAL procedure). Modeling and feature ranking are computed by the internal OFS-M procedure (e.g. a SVM for each E-RFE or RFE step). Within OFS-M, the procedure ONF is designed to estimated the optimal number of features at the saturation of a Zip's law.