Abstract— Imbalanced data set is a common problem when one set of classes has a significant advantage over the other. Two principle approaches for resampling are reducing the number of samples of the majority class (undersampling) or increasing the number of samples of the minority class (oversampling). This research uses a Generative Adversarial Network (GAN) to generate new samples for minority image classes. The key issue is how to generate good samples, i.e. that get as many features as possible of the minority class, and how to choose best samples, in the generated images, to add to minority classes. This research proposes an approach in which the Inception Score (IS) is used while training GAN to determine a good neural network configuration. Then, the samples being generated from the network with this configuration have the highest probability (at some threshold) belonging to a minority class calculated during the IS calculation will be considered as good samples and added to that class. We will simulate the imbalance in the F – MNIST data set; use WGAN with Gradient Penalty (WGAN – GP) to generate new samples for resampling. The test with CNN shows that after resampling the accuracy improved by 4,22% on minority classes.

Keywords: GAN, WGAN – GP, Inception Score, imbalanced data.