A supervised orthogonal discriminant projection for tumor classification using gene expression data

An important application of gene expression data is tumor classification. Dimensionality reduction is a key step of tumor classification, as gene expression data is of high dimensionality and small sample size (SSS) and it contains a large number of redundant genes irrelevant to tumor phenotypes. Manifold learning is an excellent tool for dimensionality reduction and it is promising for gene expression data analysis. In this paper, an improved supervised orthogonal discriminant projection (SODP) is proposed for tumor classification. In SODP, an effective weight measurement between two nodes of the weight graph is designed according to both sample class information and local information. With the novel measurement, SODP can maximize the weighted difference between the non-local scatter and the local scatter, on the basis of locality preserving. The experimental results with five public tumor datasets demonstrate that the proposed SODP is quite efficient and feasible for tumor classification.