Predicting the influence of additional training data on classification performance for imbalanced data

Abstract

It is desirable to predict the influence of additional training data on classification performance because the generation of samples is often costly. Current methods can only predict performance as measured by accuracy, which is not suitable if one class is much rarer than another. We propose an approach which is able to also predict other measures such as G-mean and F-measure, which are used in cases of imbalanced data. We show that our method leads to more correct decisions whether to generate more training samples or not using a highly imbalanced realworld dataset of scanning electron microscopy images of nanoparticles.