Feature set enhancement via hierarchical clustering for microarray classification

A new method for gene expression classification is proposed in this paper. In a first step, the original feature set is enriched by including new features, called metagenes, produced via hierarchical clustering. In a second step, a reliable classifier is built from a wrapper feature selection process. The selection relies on two criteria: the classical classification error rate and a new reliability measure. As a result, a classifier with good predictive ability using as few features as possible to reduce the risk of overfitting is obtained. This method has been tested on three public cancer datasets: leukemia, lymphoma and colon. The proposed method has obtained interesting classification results and the experiments have confirmed the utility of both metagenes and feature ranking criterion to improve the final classifier.