Sequential Oversampling for Fraud Detection: Leveraging Generative Adversarial Networks and Continual Learning Approach in Imbalanced Data Streams

Real-world fraud detection faces compounding challenges of extreme class imbalance and concept drift in evolving data streams. Existing oversampling methods, including SMOTE and GAN variants, are ill-suited for sequential learning due to their lack of temporal adaptability and memory retention. We propose Sequential GAN Oversampling with Continual Learning (SGO-CL) , a hybrid framework integrating conditional WGAN-GP with Elastic Weight Consolidation (EWC) to generate realistic minority samples across time without catastrophic forgetting. Unlike conventional GAN-based augmenters retrained per chunk, our approach incrementally fine-tunes a single generator under Fisher-informed regularization, ensuring consistency with historical fraud patterns while adapting to novel distributions. Empirical results on credit card fraud detection show that SGO-CL achieves state-of-the-art F1-scores across KNN, RF, and XGB, with up to 39.83%, 14.19%, and 25.66% gains consecutively for each model over non- augmented baselines and shows a far superior performance based on the F1-score to both SMOTE and generative baselines across the models. By coupling memory-aware generation with streaming adaptability, SGO-CL offers a scalable and model-agnostic solution for imbalanced data classification under distributional shift.