Mitigating Class Imbalance in U.S. Financial Security: A Latent Feature Augmentation Approach for High-Purity Fraud Detection

Abstract

Credit card fraud detection is challenged by extreme class imbalance (< 0.2% fraud) and non-linear fraud patterns, causing traditional supervised models to suffer from reduced sensitivity and poor interpretability. This paper proposes a hybrid model combining a sparse Deep Autoencoder (15-node bottleneck) with Random Forest classification via feature augmentation—concatenating raw features with latent representations and reconstruction errors. Validation on 85,442 transactions yields PR-AUC of 0.6942, precision of 0.9412 ± 0.0318, and recall of 0.7968 ± 0.0354 across 5-fold cross-validation. While performance is statistically comparable to standard Random Forest (p > 0.05), the hybrid model offers enhanced interpretability: the Anomaly Score ranks as the second most important predictor, and t-SNE reveals distinct geometric separation of fraud clusters. These findings demonstrate that unsupervised feature augmentation provides a viable, interpretable solution for real-time fraud monitoring.