Electronic Thesis/Dissertation
 

AI/ML-Driven Methods for Fraud Anomaly Detection

Open Access Deposited

Cost-sensitive evaluation, optimizing decision thresholds through a loss function α·FP + β·FN to reduce operational loss rather than maximize metrics. Synthetic data augmentation using Conditional Tabular Generative Adversarial Networks (CTGAN) to estimate minority-class representation under auditable pre-merge gates that govern real-data precision at t^* and Cost at t^*. Anomaly-feature fusion, combining unsupervised Isolation Forest and Autoencoder signals into supervised models to advance cost efficiency while maintaining recall. Explainable validation, employing SHAP-based interpretability on real and synthetic validation data to support regulator-ready transparency while protecting confidentiality. Experiments on a publicly available credit-card transaction dataset used a fixed, leakage-controlled 60/20/20 split. Results show that cost-aware thresholding substantially reduced total expected loss compared with metric-based operating points. Privacy-safe SHAP parity confirmed that synthetic validation can reproduce real-data feature-importance structure (ρ ≈ 0.96)

and anomaly-feature fusion improved cost efficiency while preserving recall. CTGAN augmentation did not improve performance on this dataset/split, a documented negative finding that informs stricter, policy-based acceptance tests for synthetic data. Overall, the method is transparent, auditable, and aligned with how fraud teams measure success in cost and workload, not just in scores. Future work extends the framework toward federated learning to enable multi-institutional collaboration without raw-data sharing, thus preserving cost efficiency, transparency, and privacy at scale.

Abstract A key challenge arising from the ever-increasing instances of financial fraud is the need to develop a fraud detection model that is cost-effective, understandable, as well as private. This Praxis research develops a cost-aware and privacy-suited method for detecting fraud in imbalanced financial transaction data.The methods combine four interdependent components

Author Language Keyword Date created Type of Work License
  • All rights reserved
Rights statement GW Unit Degree Advisor Committee Member(s) Persistent URL

Notice to Authors

If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.

Thumbnail Title Date Uploaded Visibility Actions
Preview of Okhankhuele_gwu_0075A_17704.pdf Okhankhuele_gwu_0075A_17704.pdf 2025-12-15 Open Access