Hybrid Graph-Ensemble Framework for E-Commerce Fraud Detection

Guggilam, Sai Rahul (2025) Hybrid Graph-Ensemble Framework for E-Commerce Fraud Detection. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (2MB) | Preview

Abstract

E-commerce platforms are losing a lot of money because of financial fraud. The current detection methods make practitioners choose between using graph-based methods to find relational patterns or using unsupervised methods to work without labelled data. This study presents a hybrid framework that combines multi-view graph construction with Isolation Forest-based anomaly detection and supervised ensemble enhancement for fraud detection on the IEEE-CIS dataset comprising 590,540 e-commerce transactions.

The methodology creates three complementary graph views that show relationships between card-email-device, card-address-BIN, and card-product-type. These views are then combined into a single graph with 16,810 nodes and 120,924 edges. Graph-based features such as multi-hop fraud exposure, fraud propagation scores, and Bayesian-smoothed entity risk ratings are extracted and integrated with conventional transaction features. The final predictions come from a hybrid ensemble that combines 30% unsupervised graph-based signals with 70% supervised gradient boosting classifiers.

From the experimental evaluation on the framework proposed above, one can determine that the framework has a ROC-AUC of about 0.8776 and a recall rate of 62.12%. This is a relative improvement on the baselines by 2.57% and is still able to learn structure and identify objects without supervision. The multi-view approach for building graphs is superior to the single-view method, and the risk scores for entities give the best unsupervised signal and constitute 70% for ranking.

The study presents a comprehensive approach that proves the feasibility of fraud detection being possible via graph ensemble methods that overcome complex design issues involved in deep learning models. Future work would be best focused on the exploration of explainability and graph dynamic processes for emerging fraud patterns.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Tomer, Vikas UNSPECIFIED
Subjects:	H Social Sciences > HG Finance H Social Sciences > HF Commerce > Electronic Commerce Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	07 Sep 2026 10:54
Last Modified:	07 Sep 2026 10:54
URI:	https://norma.ncirl.ie/id/eprint/9861

Actions (login required)

View Item