Design and Analysis of Ensemble Classifier for Gene Expression Data of Cancer

Abstract

Gene expression levels are important for disease, such as, Cancer diagnosis. This paper proposed a SVM-based ensemble classifier to classify the control and cancer groups based on gene expression levels from microarray data. A combinational Recursive Feature Elimination in conjunction with the Adaboost algorithm was developed to select significant features and design the proper classifier. The method is applied to microarray data of cancer patients, and the results show improvements on the success rate. By AUC calculation, the SVM-based ensemble classifier shows predominate performance. Furthermore, the characteristics and different effect issues to classification performance is discussed. If a single SVM can obtain satisfactory classification performance, an ensemble SVM is hardly capable to improve it. Otherwise, an ensemble of SVM is superior to the best single SVM. We also investigated the effect of kernel functions, feature selections and type of classifiers on the classification.