BAMTGAN: A Balanced Augmentation Technique for Tabular Data

This paper presents BAMTGAN, a novel data augmentation technique that addresses the class imbalance problem and prevents mode collapse by utilizing a modified DCGAN model and a new similarity loss to generate diverse and realistic tabular data. BAMTGAN encodes each column to produce a feature map for each record, which is then converted back to its original tabular form an intermediate image format. Experimental results demonstrate that BAMTGAN provides a more substantial improvement in developing high-quality predictive models than existing augmentation methods. Github: https://github.com/uos-dmlab/Structured-Data-Augmentation.git