Conditional Wasserstein GAN-based Oversampling of Tabular Data for Imbalanced Learning

Class imbalance is a common problem in supervised learning and impedes the\npredictive performance of classification models. Popular countermeasures\ninclude oversampling the minority class. Standard methods like SMOTE rely on\nfinding nearest neighbours and linear interpolations which are problematic in\ncase of high-dimensional, complex data distributions. Generative Adversarial\nNetworks (GANs) have been proposed as an alternative method for generating\nartificial minority examples as they can model complex distributions. However,\nprior research on GAN-based oversampling does not incorporate recent\nadvancements from the literature on generating realistic tabular data with\nGANs. Previous studies also focus on numerical variables whereas categorical\nfeatures are common in many business applications of classification methods\nsuch as credit scoring. The paper propoes an oversampling method based on a\nconditional Wasserstein GAN that can effectively model tabular datasets with\nnumerical and categorical variables and pays special attention to the\ndown-stream classification task through an auxiliary classifier loss. We\nbenchmark our method against standard oversampling methods and the imbalanced\nbaseline on seven real-world datasets. Empirical results evidence the\ncompetitiveness of GAN-based oversampling.\n