The area under receiver operating characteristic curve (AUC) is one of the widely used metrics for measuring imbalanced data classification results. Designing multi-objective evolutionary algorithms for AUC maximization problem has attracted much attention of researchers recently. However, most of these methods either search the Pareto front directly, or perform tailored convex hull search for AUC maximization. None of them take the advantage of multi-level knee points found in the process of evolution for AUC maximization. To this end, this paper proposes a multi-level knee point based multi-objective evolutionary algorithm (named MKnEA-AUC) for AUC maximization on the basis of a recently developed knee point driven evolutionary algorithm for multi/many-objective optimization. In MKnEA-AUC, an adaptive clustering strategy is proposed for automatically determining the knee points on the current population. By utilizing the preference of found knee points, the evolution of the population can converge quickly. We verify the effectiveness of the proposed algorithm MKnEA-AUC on 13 widely used benchmark data sets and the experimental results demonstrate that MKnEA-AUC is superior over the state-of-the-art algorithms for AUC maximization.

中文翻译：

---

基于多级拐点的多目标进化算法，用于AUC最大化

在一个REA ü的nDer接受者操作特征Çurve（AUC）是衡量不平衡数据分类结果的广泛使用的指标之一。设计用于AUC最大化问题的多目标进化算法最近引起了研究人员的广泛关注。但是，这些方法中的大多数要么直接搜索Pareto前沿，要么执行量身定制的凸包搜索以实现AUC最大化。它们都没有利用AUC最大化的进化过程中发现的多级拐点的优势。为此，本文基于最近开发的用于多/多目标优化的拐点驱动进化算法，提出了一种基于多层拐点的多目标进化算法（称为MKnEA-AUC）用于AUC最大化。在MKnEA-AUC中 提出了一种自适应聚类策略，用于自动确定当前人群的拐点。通过利用发现的拐点的偏好，种群的进化可以迅速收敛。我们在13个广泛使用的基准数据集上验证了所提算法MKnEA-AUC的有效性，实验结果表明，MKnEA-AUC优于AUC最大化的最新算法。