Training neural network classifiers on multi-class imbalanced data via competitive coevolution

In the proposed coevolutionary scheme, a population of classifiers (the predators) is pitted against a population of training patterns (the prey). The classifiers learn to correctly identify (capture) the training patterns, while the training patterns are selected based on their misclassification rate (ability to avoid capture). This mechanism of intelligent data undersampling was tested on twelve multi-class learning problems, corresponding to four different feature spaces each sampled with three different levels of severity of class imbalance. Two coevolutionary algorithms were evaluated, tasked with training a population of multi-layer perceptron classifiers. Their performance was compared to the performance of a standard evolutionary algorithm and the backpropagation rule, each used on the raw imbalanced data, or distributions rebalanced via data undersampling or oversampling. The tests highlighted the superior accuracy of the proposed approach, thanks to its competitive recognition rates on the minority class and top recognition rates on the other classes. The computational overheads of the proposed approach were modest. Undersampling boosted the classification accuracy on the underrepresented classes, but degraded the learning results on the remaining classes, while oversampling created large data sets that considerably increased the computational cost of the neural network training procedure. The tests indicated that the proposed approach is a promising solution for many of today’s classification problems involving large data sets, complex data structures like images and point clouds, and deep neural network architectures. • Competitive coevolution was demonstrated to solve imbalanced class learning problems. • Two predator–prey ANN training algorithms were tested on 12 purpose-built benchmarks. • Competitive coevolution outperformed standard evolutionary ANN learning and BP. • Competitive coevolution enabled more accurate learning than under-sampling and over-sampling. • Competitive coevolution was very economical in terms of computational overheads.