Variational Autoencoders: Tackling Imbalanced Data through Generative Modeling

With the advancement of artificial intelligence models, the challenge of working with real-world data has become increasingly important in fields such as computer vision, autonomous vehicles, medical and forecasting. Among them, the issue of imbalanced data in real datasets is very crucial. While the capabilities of Variational Autoencoders (VAEs) have been introduced in various fields, this study is the first to leverage VAE as a powerful tool to tackle the challenge of imbalanced data in real-world scenarios. In this study, a cluster-based oversampling model based on VAE is introduced. Following the formation of the latent space during training and learning the data distribution, VAE is employed to generate synthetic data and balance the dataset based on the learned parameters. This model has been implemented on a real-world imbalanced dataset that we have developed, and it is shown it well outperforms traditional oversampling methods and Generative Adversarial Network (GAN) in all evaluation rubrics. Additionally, this model has significantly lower computational costs. Given these advantages, the proposed model is very promising to be harnessed in different fields that are encountering imbalanced data challenges.