A multivariate algorithm for gene selection based on the nearest neighbor probability

Since DNA-microarray datasets include a very high number of genes, in the last few years researchers have focused their attention on algorithms capable to select a subset of the input features which can classify (e.g. ill/healthy) the patterns with a sufficient level of accuracy. Some of the methods proposed to solve this problem are based on Recursive Feature Addition (RFA). According to this approach, at each iteration the gene which maximizes a proper discriminant function \phi is selected; then \phi is updated in conformity with the performed choice. In this paper an RFA method for gene selection based on nearest neighbor probability, named NN-RFA, is described and tested on some artificial datasets simulating the behavior of human tissues. The results of such simulations show the ability of NN-RFA of retrieving a correct subset of genes for the problem at hand.