STACK-GFD: Stacked Ensemble Learning for Graph-Based Fraud Detection

Overview

STACK-GFD is a two-level stacked ensemble combining four GNN-based anomaly detectors with an XGBoost meta-learner for fraud detection across multiple graph domains.

Results

Dataset	AUC-ROC	Improvement
Amazon	95.08%	+19.78%
DGraph-Fin	83.50%	+40.81%
MulDiGraph	99.70%	+5.39%

Requirements

See requirements.txt

Datasets

• Amazon: Available via PyGOD BOND benchmark
• DGraph-Fin: https://dgraph.ecs.soton.ac.uk/
• MulDiGraph: http://xblock.pro/ethereum/

How to Run

1. Install requirements: pip install -r requirements.txt
2. Run notebook