Application of graph regularized non-negative matrix factorization in characteristic gene selection

Abstract

Nonnegative matrix factorization (NMF) has become a popular method and widely used in many fields, for the reason that NMF algorithm can deal with many high dimension, non-negative problems. However, in real gene expression data applications, we often have to deal with the geometric structure problems. Thus a Graph Regularized version of NMF is needed. In this paper, we propose a Graph Regularized Non-negative Matrix Factorization (GRNMF) with emphasizing graph regularized on error function to extract characteristic gene set. This method considers the samples in low-dimensional manifold which embedded in a high-dimensional ambient space, and reveals the data geometric structure embedded in the original data. Experiment results on tumor datasets and plants gene expression data demonstrate that our GRNMF model can extract more differential genes than other existing state-of-the-art methods.