Balancing Act: MDGAN for Imbalanced Tabular Data Synthesis

Addressing the persistent challenge of learning from imbalanced datasets is crucial in advancing machine learning applications. Standard machine learning algorithms typically assume that the input data is balanced, and they often struggle to effectively learn the distribution of minority class data when dealing with imbalanced data. To address this, our study designed an improved Generative Adversarial Networks (GANs) model, named MDGAN, for tabular sample synthesis to augment samples and balance the data distribution. MDGAN employs a multi-generator and multi-discriminator structure to capture non-connected subspace manifolds, thereby better fitting the complete data distribution. To enhance the diversity among the multiple generators, an exclusive loss among generators was designed, ensuring that each generator produces data of different modalities. Additionally, a contrastive loss was introduced to ensure that the generated samples better fit the minority class distribution and are separated from the majority class distribution, preventing blurred classification boundaries. Qualitative and quantitative tests were conducted on 25 real datasets, and the experimental results indicate that MDGAN outperforms traditional classical models and current advanced oversampling models.