Temporal-Aware Graph Attention Network with XGBoost Ensemble for Blockchain Transaction Fraud Detection

Abstract

Blockchain transaction fraud detection is a critical challenge as cryptocurrency adoption grows rapidly. Graph neural networks (GNNs) and temporal models show potential, but key questions persist about their advantages over tabular baselines, stability in realistic class distributions, and the true predictive value of temporal patterns. This paper presents a systematic empirical study across two real-world blockchain datasets-the Bitcoin Elliptic dataset and an Ethereum phishing dataset-comparing ten models including classical machine learning (ML) methods, five GNN architectures, a Temporal-Aware Graph Attention Network (ATGAT), and a Hybrid ATGAT+XGBoost ensemble. Experiments are conducted under both balanced (50:50) and imbalanced settings using strict temporal splits to prevent future leakage. Our results reveal that XGBoost remains a formidable baseline, achieving $\mathrm{F} {1}={0.6889}$ and $\text{ROC-AUC}=\{0.8731\}$ on Bitcoin, while GNN-only approaches consistently underperform under temporal evaluation. The Hybrid model achieves competitive performance $(\mathrm{F} {1} {=} {0.6816}$ on Bitcoin; ROC-AUC ={0.9874} on Ethereum) and provides more robust generalisation than pure GNN models, though it does not clearly surpass standalone XGBoost on Bitcoin. Temporal fraud rate analysis confirms that fraud dynamics evolve over time, motivating temporal-aware architectures even when their measured gains are modest.