A new maximum-relevance criterion for significant gene selection

Abstract

Gene (feature) selection has been an active research area in microarray analysis. Max-Relevance is one of the criteria which has been broadly used to find features largely correlated to the target class. However, most approximation methods for Max-Relevance do not consider joint effect of features on the target class. We propose a new MaxRelevance criterion which combines the collective impact of the most expressive features in Emerging Patterns (EPs) and some popular independent criteria such as t-test and symmetrical uncertainty. The main benefit of this criterion is that by capturing the joint effect of features using EPs algorithm, it finds the most discriminative features in a broader scope. Experiment results clearly demonstrate that our feature sets improve the class prediction comparing to other feature selections. © Springer-Verlag Berlin Heidelberg 2006.