Adaptive Ensemble Learning Framework with SHAP-Based Feature Optimization for Financial Anomaly Detection

Abstract

Financial fraud detection remains a critical challenge in digital banking infrastructure, requiring sophisticated algorithmic approaches that balance accuracy with interpretability. This paper presents an adaptive ensemble learning framework that integrates XGBoost, LightGBM, and CatBoost algorithms with SHAP-based feature optimization to enhance anomaly detection capabilities in financial transactions. The proposed framework addresses class imbalance through SMOTE-ENN hybrid sampling while maintaining computational efficiency for real-time applications. Our methodology incorporates dynamic feature selection using SHAP values, achieving global interpretability essential for regulatory compliance. Experimental evaluation on benchmark datasets demonstrates superior performance with 97.3% AUC-PR, outperforming traditional isolation forest and single gradient boosting approaches by by 12.6 percentage points and the best single gradient-boosting baseline (0.924) by 4.9 points (≈5.3% relative), respectively. The framework's interpretability analysis reveals critical risk factors through SHAP visualizations, providing actionable insights for fraud prevention strategies while maintaining sub-second inference latency for production deployment.