Gene Selection and Visualization Based on Sparse Maximal Margin Features

Gene selection with interpretation is an important problem in the bioinformatics field. A novel approach called sparse maximal margin features is proposed in this paper for gene subsets selection and visualization. Through transforming an dense eigenvalue decomposition problem into the Elastic-Net regularized sparse regression framework, we introduce sparsity constraint into the coefficients, which is useful to enhance the interpretability of important variables. Moreover, the new method can simultaneously maximize between-class scatter while minimize within-class scatter, and avoid the small sample size problem. The experimental results from gene expression data show that, our method is helpful to select discriminant genes and then provide important foundations for cancer diagnosis.