Maximizing receiver operating characteristics convex hull via dynamic reference point-based multi-objective evolutionary algorithm

The receiver operating characteristic convex hull (ROCCH) is a popular technique for analyzing the performance of classifiers, which is particularly effective for the tasks with unbalanced data distribution. Although maximization of ROCCH can be tackled as a bi-objective optimization problem, existing multi-objective evolutionary algorithms (MOEAs) encounter difficulties in obtaining an ROCCH, since ROCCH is always convex but the Pareto front obtained by MOEAs may be concave. To address the issue, in this paper, a dynamic reference point-based MOEA, namely DR-MOEA is proposed for maximizing ROCCH performance. Specifically, in DR-MOEA, a reference point-based sorting is suggested, where the solutions are sorted by their distances to the reference points instead of Pareto dominance. Hence an ROCCH rather than a Pareto front is expected to be obtained. In addition, a reference point adaptation strategy is also designed, with which the reference points are dynamically adjusted during the evolutionary process, and the performance of DR-MOEA is further enhanced. Empirical studies are conducted by comparing the proposed algorithm with several state-of-the-arts on different data sets. Experimental results demonstrate the superiority of DR-MOEA over the comparison methods in solving the ROCCH maximization problem. (C) 2019 Elsevier B.V. All rights reserved.