Implementation of the cWGAN-based oversampling method. Fits a conditional Wasserstein GAN with Gradient Penalty and an auxiliary classifier loss to a tabular dataset with categorical and numerical attributes. The fitted cWGAN model can than be used to resample an imbalanced training set. Currently only supports binary classification.
Our implementation was initially based on [1] and also drew upon various WGANGP pytorch implementations such as [2] [3] [4] .
The datasets used by our evaluation are not included in this repository but are linked to in dataloader.py. At the time of writing, all the datasets are publicly available.