Genetic programming for image classification with unbalanced data

Image classification methods using unbalanced data can produce results with a performance bias. If the class representing important objects-of-interest is in the minority class, learning methods can produce the deceptive appearance of ¿good looking¿ results while recognition ability on the important minority class can be poor. This paper develops and compares two genetic programming (GP) methods for image classification problems with class imbalance. The first focuses on adapting the fitness function in GP to evolve classifiers with good individual class accuracy. The second uses a multi-objective approach to simultaneously evolve a set of classifiers along the trade-off surface representing minority and majority class accuracies. Evaluating our GP methods on two benchmark binary image classification problems with class imbalance, our results show that good solutions were evolved using both GP methods.