An Explainable AI (XAI) Framework for Anomaly Detection in Financial Transactions

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The rapid digitalization of global financial infrastructure has catalyzed a parallel evolution in adversarial transaction patterns, rendering static rule-based detection systems obsolete. while Deep Neural Networks (DNNs) offer superior capacity for modeling non-linear dependencies in high-dimensional banking data, their deployment is severely constrained by the "black-box" opacity problem. This lack of transparency conflicts with rigorous regulatory mandates such as GDPR and Basel III, which enforce the right to explainability in automated decision-making. This manuscript introduces D-PEAR (Dual-Phase Explainable Anomaly Reconnaissance), a hybrid architecture fusing Variational Autoencoders (VAE) with Attention-augmented Long Short-Term Memory (LSTM) networks. Unlike traditional approaches that treat detection and explanation as separate silos, D-PEAR integrates latent space probabilistic modeling with a post-hoc interpretability engine utilizing SHAP (Shapley Additive exPlanations) and LIME. Benchmarking against Isolation Forests, Local Outlier Factors (LOF), and standard Autoencoders on a heavily imbalanced credit card fraud dataset reveals that D-PEAR achieves a state-of-the-art Area Under the Precision-Recall Curve (AUPRC) of 0.842. Furthermore, the framework provides granular, instance-level attribution maps, successfully bridging the dichotomy between predictive efficacy and regulatory auditability.