Discriminative feature generation for classification of imbalanced data

The data imbalance problem is a frequent bottleneck in the classification performance of neural networks. In this paper, we propose novel supervised discriminative feature generation (DFG) method for minority class dataset. DFG based on modified structure generative adversarial network consisting four independent networks: generator, discriminator, extractor, and classifier. To augment selected features by adopting an attention mechanism, generator class-imbalanced target task trained, extractor classifier are regularized using pre-trained from large source data. experimental results show that enhances augmentation label-preserved diverse features, significantly improved task. model can contribute greatly to development methods through methods.