- Conference Article
- 10.1109/smap66932.2025.00029
Explainable Generative AI Models for Anomaly Detection in Transactional Finance Data
- Nov 27, 2025
- Sowjanya Karri
Anomaly detection in financial transactions is a critical component in combating fraud, ensuring regulatory compliance, and safeguarding user trust. With the increasing complexity of online transaction system, traditional rule-based methods are often insufficient in identifying subtle or evolving fraud patterns. Deep generative models, particularly Variational Autoencoders (VAEs), have emerged as powerful tools for unsupervised anomaly detection in high-dimensional financial datasets. However, these methods often suffer from a lack of interpretability, limiting their adoption in transparent and regulated financial environments. To address this gap, we propose an explainable generative AI framework that integrates attention-enhanced VAE architectures with post-hoc interpretability tools such as SHAP. Our model not only detects anomalous behavior but also offers transaction-level attribution that help human analysts understand the reasoning behind outlier classification. Evaluated on the IEEE-CIS Fraud Detection dataset, the model demonstrates high performance in identifying rare fraudulent patterns while maintaining a balance between detection accuracy and interpretability. The proposed methodology represents a significant step towards the practical deployment of explainable generative AI in financial anomaly detection workflow.
Read more