Multi-objective evolutionary algorithm for optimizing the partial area under the ROC curve

摘要

The measure called partial area under the curve (pAUC) has attracted increasing interest in recent years due to the wide range of real-world applications it can be used for, such as biomarker selection and pedestrian detection. Compared with AUC, pAUC only acts upon a partial (instead of a full) area under the ROC curve. In this paper, a multi-objective evolutionary algorithm termed MOPA is proposed to optimize the pAUC in an arbitrary false positive range. Within this range, two important components are proposed to focus on a particular region within the AUC. First, a new metric (K-FPR), is proposed that is created by considering the partial range of the false positive rate (FPR). This is combined with the true positive rate (TPR) to provide the two optimized objectives of MOPA. Second, a preference-based multi-objective evolutionary algorithm is developed within the framework of a recently proposed algorithm (AR-MOEA), by which the search is concentrated on a partial area under the ROC curve. Numerical experiments on different data sets demonstrate the competitiveness of the proposed method in comparison with state-of-the-art algorithms.