A Class-information-based SNMF method for selecting characteristic genes

The significant advantage of sparse methods is to reduce the complicacy of genes expression data, which makes them easier to understand and interpret. In this paper, we propose a novel Class-information-based Sparse Non-negative Matrix Factorization (CISNMF) method which introduces the class information by the total scatter matrix. Firstly, the total scatter matrix is obtained via combining the between-class and within-class scatter matrices. Secondly, a new data matrix is constructed via singular values and left singular vectors which can be obtained via decomposing the total scatter matrix. Finally, we decompose the new data matrix by using sparse Non-negative Matrix Factorization and extract characteristic genes. In the end, results on gene expression data sets show that our method can extract more characteristic genes in response to abiotic stresses than conventional gene selection methods.