Convex-hull based robust evolutionary optimization approach for ROC maximization under label noise

Convex-hull based receiver operating characteristic (ROC) maximization has become a hot research topic due to its significance in the study of imbalance binary classification. Recently, a series of multi objective evolutionary algorithms (MOEAs) have been proposed to maximize ROC convex hull by regarding it as a multi-objective optimization problem. However, in real applications, the ubiquitous label noises degrade the performance of the existing MOEAs. To address this issue, in this paper, we propose a robust evolutionary optimization approach, named REO, to enhance the robustness of the existing MOEAs. In the proposed approach, we firstly design a distance-based samples selection method to extract a "clean"data subset, aiming to obtain an ideal individual. Second, with the ideal individual, a problem-oriented two-stage adaptive updating strategy is designed to guide the population evolution and enhance the robustness of MOEAs. Specifically, in the first stage, based on the achieved ideal individual, a bi-level evolution direction is constructed to provide the guidance for the evolution of population. In the second stage, we utilize the cosine similarity to assign different step sizes to adaptively update the inferior individuals. Experimental results on 19 complicated datasets with different noise levels show that the proposed REO approach can effectively enhance the robustness of the existing MOEAs for ROC convex hull maximization under the label noises.& COPY; 2023 Elsevier B.V. All rights reserved.