Advanced R-GAN: Generating anomaly data for improved detection in imbalanced datasets using regularized generative adversarial networks

The high prevalence of fraud in contemporary financial transactions necessitates advanced anomaly detection systems to address the significant imbalance between legitimate and anomalous transactions in real-time datasets. To address this issue, our study introduces a novel approach, the regularized generative adversarial network (R-GAN). Diverging from conventional resampling techniques and typical generative adversarial network (GAN) architectures, R-GAN incorporates spectral normalization for the STGAN (short for spectral normalization for GAN) generator framework, which enhances it with a similarity measure loss to improve the authenticity of the generated data. The discriminator is meticulously designed, leveraging the CELU (short for continuously differentiable exponential linear unit) activation for optimal feature extraction, ensuring diverse and representative sample generation. To ensure fairness and validate the effectiveness of our data generation process, we used PyCaret's automated machine learning framework to rigorously test different machine learning models, ultimately identifying the light gradient boosting machine as the most effective. To add transparency to our system, we applied Shapley additive explanations (SHAP), providing clear insights into the decisions made by our explainable artificial intelligence-driven model. This approach ensures high-fidelity anomaly detection in real-time environments and continuously refines through SHAP insights, significantly addressing imbalanced datasets across various applications.