Improved mutual information-based gene selection with fuzzy rough sets

Feature selection is an essential step to perform cancer classification with DNA microarrays. Rough set theory has already been successfully applied to gene selection. To avoid losing information by discretization of continuous gene expression data in rough set approach, the fuzzy rough set theory has been applied to gene selection. A fuzzy rough attribute reduction algorithm based on mutual information has been proposed. However, the cost of computation of the algorithm is too high to be carried out if the number of the selected genes is large. In this paper, an approximate replacement of the mutual information, from both maximum relevance and maximum significance is raised. The novel method improves the efficiency and decreases the complexity of the classical algorithm. Extensive experiments are conducted on three public gene expression datasets and the experimental results confirm the efficiency and effectiveness of the proposed algorithm. © 2005 by Binary Information Press.