Multi-objective evolution of artificial neural networks in multi-class medical diagnosis problems with class imbalance.

Abstract

This paper proposes a novel multi-objective optimisation approach to solving both the problem of finding good structural and parametric choices in an ANN and the problem of training a classifier with a heavily skewed data set. The state-of-the-art CMA-PAES-HAGA multi-objective evolutionary algorithm [41] is used to simultaneously optimise the structure, weights, and biases of a population of ANNs with respect to not only the overall classification accuracy, but the classification accuracies of each individual target class. The effectiveness of this approach is then demonstrated on a real-world multi-class problem in medical diagnosis (classification of fetal cardiotocograms) where more than 75% of the data belongs to the majority class and the rest to two other minority classes. The optimised ANN is shown to significantly outperform a standard feed-forward ANN with respect to minority class recognition at the cost of slightly worse performance in terms of overall classification accuracy.