Improving performance of gene selection by unsupervised learning

Selection of significant genes via expression profiles is an important problem in microarray experiments for diseases classification and prediction. Genes of interest are typically selected by a statistical significance test and the top ranked genes were used. A problem with this approach is that many of these genes are highly correlated. For classification purposes it required to have distinct but still highly informative genes. In this paper, we proposed an unsupervised feature selection algorithm to resolve this problem. The method retrieves groups of similar genes by measuring similarity between them whereby redundancy therein is removed. This does not need any search and therefore, is fast. Real biological data experiments have shown that this approach will significantly improve existing classifiers.