Evolutionary fuzzy classifiers for imbalanced datasets: An experimental comparison

In this paper, we compare three state-of-the-art evolutionary fuzzy classifiers (EFCs) for imbalanced datasets. The first EFC performs an evolutionary data base learning with an embedded rule base generation. The second EFC builds a hierarchical fuzzy rule-based classifier (FRBC): first, a genetic programming algorithm is used to learn the rule base and then a post-process, which includes a genetic rule selection and a membership function parameters tuning, is applied to the generated FRBC. The third EFC is an extension of a multi-objective evolutionary learning scheme we have recently proposed: the rule base and the membership function parameters of a set of FRBCs are concurrently learned by optimizing the sensitivity, the specificity and the complexity. By performing non-parametric statistical tests, we show that, without re-balancing the training set, the third EFC outperforms, in terms of area under the ROC curve, the other comparison approaches.