Research on modeling of the imbalanced fraudulent transaction detection problem based on embedding-aware conditional GAN

Detecting fraudulent transactions in structured financial data presents significant challenges due to multimodal, non-Gaussian continuous variables, mixed-type features, and severe class imbalance. To address these issues, we propose an Embedding-Aware Conditional Generative Adversarial Network (EAC-GAN), which incorporates trainable label embeddings into both the generator and discriminator to enable semantically controlled synthesis of minority-class samples. In addition to adversarial training, EAC-GAN introduces an auxiliary classification objective, forming a joint optimization strategy that improves the fidelity and class consistency of generated data, especially for underrepresented classes. Experiments conducted on a real-world credit card dataset demonstrate that EAC-GAN achieves stable convergence even with limited labeled data. When combined with LightGBM classifiers, the synthetic samples generated by EAC-GAN significantly enhance fraud detection performance, yielding a precision of 96.8%, an AUC of 96.38%, an AUPRC of 83.89%, and an MCC of 88.94%. Furthermore, dimensionality reduction using Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE) reveals that the generated samples closely align with the real data distribution and exhibit clear class separability in the latent space. These results underscore the effectiveness of EAC-GAN in synthesizing high-quality minority-class samples and improving downstream fraud detection, outperforming traditional oversampling techniques and baseline generative models.