Leveraging Artificial Intelligence for Enhanced Data Generation in Addressing Imbalance in Binary Classification System

This paper delves into the challenges of binary classification using imbalanced datasets, particularly when instances of interest are infrequent. It explores a comprehensive approach that integrates Synthetic Minority Over-sampling Technique (SMOTE), Generative Adversarial Networks (GANs), and Variational Autoencoders (VAEs) to enhance classification outcomes. Traditional classification models tend to favor the majority class, while the impact of imbalanced misclassification costs is often overlooked. The integration of SMOTE, GANs, and VAEs in binary classification, or SMOTE-GAN-VAE, addresses these challenges by generating synthetic instances, refining data representations, and capturing latent features. To evaluate the effectiveness of various data generation methods, a credit card fraud dataset is used. The performance metrics considered include F0.5-score, F1-score, and F2-score, which account for both precision and recall. The results indicate that SMOTE-GAN-VAE outperforms individual methods, such as SMOTE, GANs, and VAEs, demonstrating its potential to enhance data representation and classification accuracy, and outperformed the β- VAE filtered approach employed in previous literature.