SVD based gene selection algorithm

Abstract

This paper proposes an unsupervised gene selection algorithm based on the singular value decomposition (SVD) to determine the most informative genes from a cancer gene expression dataset. These genes are important for many tasks including cancer clustering and classification, data compression, and samples characterization. The proposed algorithm is designed by making use of the SVD's clustering capability to find the natural groupings of the genes. The most informative genes are then determined by selecting the closest genes to the corresponding cluster's centers. These genes are then used to construct a new (pruned) dataset of the same samples but with less dimensionality. The experimental results using some standard datasets in cancer research show that the proposed algorithm can reliably improve performances of the SVD and kmeans algorithm in cancer clustering tasks.