Explainable Artificial Intelligence for Anomaly Detection in Banking Transactions: An Internal Audit Perspective

Explainability builds trust in automated fraud detection: The paper demonstrates that when AI systems provide feature-level explanations for why a transaction was flagged as suspicious, audit professionals report measurably improved confidence in the system's decisions—addressing a critical gap where conventional automated systems offer no justification for their outputs.

High accuracy with fewer false alarms: The framework achieved 0.91 precision and 0.88 recall on banking data, meaning it correctly identifies suspicious transactions while reducing the false-positive rates that plague traditional rule-based approaches, potentially saving compliance teams significant time investigating legitimate transactions.

Making AI accessible to non-technical users matters: The researchers built a dashboard interface using Streamlit specifically designed for audit professionals without machine learning expertise, recognizing that responsible AI deployment requires tools that enable human decision-makers to understand and act on algorithmic outputs, not just receive them.

Game theory provides a principled foundation for explanations: The use of SHAP (based on cooperative game theory) grounds the explanations in a mathematically rigorous framework rather than ad-hoc reasoning, which is important for regulated industries where audit trails and decision justification must withstand regulatory scrutiny.

Unsupervised learning suits compliance workflows: By using Isolation Forest for anomaly detection rather than supervised methods, the system can flag novel or unusual transaction patterns without requiring labeled historical fraud data, making it more practical for real-world deployment where comprehensive fraud labels are often unavailable.