Hybrid Explainable Autoencoders for Credit Card Fraud Detection: Integrating Deep Latent Representations with Model-Agnostic Interpretability

Credit card fraud remains a persistent challenge in the financial sector, exacerbated by the evolving sophistication of fraudulent behavior and the scarcity of labeled data. This study proposes a hybrid anomaly detection framework that combines a deep autoencoder with SHapley Additive Explanations (SHAP) to improve interpretability. The model is trained in an unsupervised manner using only legitimate transactions and leverages reconstruction error to detect anomalous patterns indicative of fraudulent activity. The approach is evaluated using a publicly available, highly imbalanced credit card transaction dataset. The autoencoder achieves a ROC AUC score of 0.9662, demonstrating strong performance in identifying fraud without reliance on labeled data. Comparative benchmarks are provided using Isolation Forest and XGBoost, with results showing that while XGBoost offers higher predictive accuracy (ROC AUC = 0.9705), it requires fully labeled datasets. The proposed method offers a competitive alternative, particularly suited for dynamic environments where labeled anomalies are limited or unavailable. The integration of SHAP allows for transaction-level explanations of anomaly scores, supporting greater transparency and trust in automated decision-making. These findings suggest that hybrid explainable autoencoders represent a viable and practical solution for real-world fraud detection systems. This study is one of the first implementations applying SHAP to interpret reconstruction-based unsupervised fraud detection.