On Oversampling via Generative Adversarial Networks under Different Data Difficulty Factors
Proceedings of the Third International Workshop on Learning with Imbalanced Domains: Theory and Applications, PMLR 154:76-89, 2021.
Over the last two decades, several approaches have been proposed to tackle the class imbalance problem which is characterized by the inability of a learner to focus on a relevant but scarcely represented class. The generation of synthetic examples to oversample the training set and thus force the learner to focus on the important cases is one of such solutions. Recently, generative adversarial networks (GANs) started to be explored as an oversampling alternative due to their capability of generating samples from an implicit distribution. Still, data difficulty factors such as class overlap, data dimensionality or sample size, and were shown to also negatively impact the learners performance under an imbalance setting. The ability of GANs to deal with the imbalance problem and other data difficulty factors has not yet been assessed. The main goal of this paper is to understand how data difficulty factors impact the performance of GANs when they are used as an oversampling method. Namely, we study the performance of conditioned GANs (CGANs) in an image dataset with controlled levels of the following data difficulty factors: sample size, data dimensionality, class overlap and imbalance ratio. We show that CGANs are effective for tackling tasks with multiple data difficulty factors, exhibiting increased gains on the most difficult tasks.