Neighborhood Rough Set Model Based Gene Selection for Multi-subtype Tumor Classification

Multi-subtype tumor diagnosis based on gene expression profiles is promising in clinical medicine application. Therefore, a great deal of research on tumor classification based on gene expression profiles has been developed, where various machine learning approaches were applied to constructing the best tumor classification model to improve the classification performance as much as possible. To achieve this goal, extracting features or finding informative genes that have good classification ability is crucial. We propose a novel gene selection approach, which adopts Kruskal-Wallis rank sum test to rank all genes and then apply an algorithm based on neighborhood rough set model to gene reduction to obtain gene subsets with fewer genes and more classification ability. Experiments on a small round blue cell tumor (SRBCT) dataset show that our approach can achieve very high classification accuracy with only three or four genes as evaluated by three classifiers: support vector machines, K-nearest neighbor and neighborhood classifier, respectively.