Microarray‐based cancer diagnosis: repeated cross‐validation‐based ensemble feature selection

Abstract

The influence of data resampling on ensemble methods, and repeated cross-validation (RCV)-based ensemble feature selection (FS) is proposed. To evaluate the proposed method, support vector machine and its extension and recursive feature elimination were used as the underlying classification and FS techniques, respectively. Experimental evaluation was performed using four microarray datasets. The results show that especially for extremely small signature sizes, increasing ensemble size increases both classification performance and the robustness of gene selection (stability) for both RCV and bootstrap (BS). However, for ensembles of the same size, RCV outperforms BS in terms of performance and especially stability. When compared to the top results obtained by two other studies in which BS is utilised, RCV performs similar or better in terms of area under the receiver operator curve and better in terms of stability.