Multi-Model Financial Fraud Detection With Explainable AI

Financial fraud continues to escalate in complexity, posing significant threats to digital financial ecosystems. This study introduces a scalable and interpretable fraud detection framework that combines unsupervised anomaly detection with generative AI-based explanation. Leveraging a unified schema from three diverse datasets—PaySim, BankSim, and a real-world credit card fraud dataset—the system applies Isolation Forest for anomaly detection in imbalanced, high-dimensional data. To address the black-box nature of machine learning, we integrate Google Gemini to generate natural language explanations for each flagged transaction. Comparative analysis with supervised models (Random Forest and XGBoost), conducted on the same datasets and under similar experimental conditions, reveals that while XGBoost delivers state-of-the-art accuracy (ROC-AUC: 0.97), the proposed Isolation Forest–Gemini ensemble offers strong unsupervised detection (F1-score: 0.85) and interpretable justifications with low infrastructure requirements. This research contributes a practical, transparent, and lightweight solution for real-time fraud detection in modern financial services.