Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Katerina Potika

Second Advisor

Robert Chun

Third Advisor

Mihir Dhirajlal Satra

Keywords

Graph Neural Networks, Fraud Detection, Stacked Ensemble Learn- ing, Anomaly Detection, XGBoost, SHAP Explainability

Abstract

Fraud detection methods have many real-world applications and are essential in industries such as finance and security. The task becomes harder due to the diverse characteristics of fraud, the complex relationships among participants, and the imbalance in the number of detected frauds. Graph modeling of such problems improves fraud prediction by considering both the underlying interaction structure and fraud-related features. We propose STACK-GFD, which leverages Deep Learning on graphs for anomaly detection and a Stacked Ensemble approach. Although graph deep learning has been used for fraud detection, we investigate the benefits of ensembling multiple models to improve predictive performance. To capture varying anomalous characteristics, the ensemble’s first level uses graph-based methods (DOMINANT, AnomalyDAE, COLA, and GAT) to calculate anomaly scores for each node. In the next and final level, a meta learner, such as XGBoost, synthesizes the anomaly scores from all models of the previous level and provides a final score. Instead of simple averaging, the model learns to optimally weight each graph-based method based on the unique characteristics of each dataset. Experimental results are performed on three datasets from different areas of fraud: (Amazon-injected benchmark) for e-commerce fraud, DGraph-Fin, a financial transaction graph, and the MulDiGraph Ethereum phishing network. The results show an improvement in the AUC-ROC metric. For ensemble explainability, we use SHAP values. Our results indicate that STACK-GFD performs well in diverse datasets.

Download

Available for download on Tuesday, May 25, 2027

Share

COinS