AUC optimization Boosting based on data rebalance

The area under the receiver operating characteristics (ROC) curve (AUC) is usually used to evaluate the classifier performance over the whole class prior probability distribution. Boosting can maximize the classification accuracy, which is not optimal under the AUC measure. An improved boosting algorithm which optimizes the AUC is proposed. By introducing data rebalance operation into boosting iterations, the optimization objective of the weak learning algorithm is transferred to the AUC instead of accuracy. Experimental results show that compared with naive boosting, the new algorithm gets better performance under the AUC measure. © 2013 Acta Automatica sinica. All rights reserved.