A hybrid approach to the problem of class imbalance

In Machine Learning classification tasks, the class imbalance problem is an important one which has received a lot of attention in the last few years. In binary classification, class imbalance occurs when there are significantly fewer examples of one class than the other. A variety of strategies have been applied to the problem with varying degrees of success. Typically previous approaches have involved attacking the problem either algorithmically or by manipulating the data in order to mitigate the imbalance. We propose a hybrid approach which combines Individualised Random Sampling(IRS) with two different fitness functions designed to improve performance on imbalanced classification problems in Genetic Programming. We investigate the efficacy of the proposed methods together with that of five different algorithmic GP solutions, two of which are taken from the recent literature. We conclude that the IRS approach combined with either average accuracy or Matthews Correlation Coefficient, delivers superior results in terms of AUC score when applied to either balanced or imbalanced datasets.