RABEM: risk-adaptive Bayesian ensemble model for fraud detection.

Area of Science:

• Computational Finance
• Machine Learning for Fraud Detection
• Data Science

Background:

• Financial fraud detection is a critical challenge in digital transactions.
• Existing methodologies require enhancement for robust performance.
• Large-scale synthetic datasets are valuable for developing and testing fraud detection models.

Purpose of the Study:

• To develop an advanced computational model for improved financial fraud detection.
• To address the limitations of current fraud detection techniques.
• To leverage synthetic financial data for robust model development.

Main Methods:

• Utilized Kaggle's Synthetic Financial Datasets (6 million transactions).
• Developed the Risk Adaptive Bayesian Ensemble Model (RABEM).
• Integrated Black-Scholes Feature Engineering, Hybrid VAE, Nyström Approximation Gaussian Process, Random Projection Tree (RPTree), Gated Recurrent Unit (GRU), and Bayesian Reliability Fusion.

Main Results:

• Achieved a high accuracy of 99.38% in fraud detection.
• Demonstrated superior performance over other approaches.
• Key metrics include MCC of 0.9788, Brier Score of 0.0061, and log loss of 0.2103.
• Top-K hit rate analysis showed 97.2% precision in identifying fraudulent transactions (972/1000).

Conclusions:

• The RABEM methodology offers high accuracy and dependability for financial fraud detection.
• The model effectively distinguishes between legitimate and fraudulent transactions.
• Future work will explore larger datasets and enhanced feature selection for improved performance.