A Ranked Subspace Learning Method for Gene Expression Data Classification

Abstract

Abstract — Microarray expression data analysis is critical for clinical treatment and biomedical research. Although many research results have been reported in literatures, it is still very difficult to learn, predict and understand such complicated gene expression data. The major difficulty arises from the inherent characteristics of such data set including extremely high dimensions (up to around 30,000 ~ 40,000) and small available data samples (in the order of hundreds or less). In this paper, we proposed a general learning methodology that can improve the classification accuracy and robustness based on the feature rank and classify ensemble; we call it Ranked Subspace (RS) method. In RS method, all the features are first ranked and sorted in a descending order according to a feature score function. Then numerous classifiers are built based on the bootstrap sampling of these feature spaces with different weights according to their importance. Finally, a combination voting mechanism is used to combine all the information from different classifiers to get the final predicted class identity labels. We presented in detail the proposed RS algorithm, and tested the classification performance on 4 major biomedical databases. Performance comparisons of the RS algorithm with those of random subspace method showed the proposed method can improve the classification accuracy.