Multi-discriminator generative adversarial networks with dynamic penalty to over-sample imbalanced credit datasets

The problem of credit risk data imbalance reduces the effectiveness of assessment models. Existing oversampling methods focus only on a partial sample of a few classes, resulting in a lack of diversity in the types of data generated. This paper proposes an innovative GAN variant called Magnify-GAN. The originality of Magnify-GAN lies in the fact that it is equipped with a primary discriminator and multiple secondary discriminators, each of which employs a different loss function. This multi-discriminator approach not only improves the learning results, but also enriches the feedback received during the training process. In addition, we integrate an innovative dynamic coefficient mechanism to enable the model to dynamically adapt to changes in data distribution. To further improve stability and address the common modal collapse problem in GAN, a gradient penalty method is embedded in the training protocol. This integrated strategy ensures that Magnify-GAN can effectively generate samples representing various minority classes within the real data. Compared to ten classical imbalanced sampling methods, Magnify-GAN demonstrates superior performance in precision, F1-score, and AUC values across six synthetic and four real-world imbalanced datasets. Ablation studies, visualized through heatmaps, reveal the complementary synergy between the core modules. Furthermore, a complexity analysis shows that Magnify-GAN offers significant performance gains with moderate increases in computational cost compared to state-of-the-art methods.