Data-Driven Risk Analytics for Blockchain Transaction Fraud: A Semi-Supervised Heterogeneous Graph Modeling Framework
Main article
Abstract
Blockchain transaction networks create a distinctive environment for risk analytics because every transfer is public, yet the economic identity, intent, and future behavior of participants remain only partially observable. This article develops a data-driven semi-supervised heterogeneous graph modeling framework for detecting transaction fraud in blockchain ecosystems. Rather than treating blockchain fraud detection as a purely technical classification problem, the study frames it as a business risk analytics problem in which alert quality, analyst workload, missed illicit flow, and explainability are jointly considered. The proposed framework converts raw transfers into a dynamic heterogeneous graph with transaction nodes, address nodes, relation-specific edges, temporal snapshots, and channel-level risk attributes. A semi-supervised learning strategy is then used to exploit scarce labels, abundant unlabeled observations, and high-confidence normal behavior. The empirical section reports a controlled numerical study calculated to common public Bitcoin anomaly-detection settings and evaluates logistic regression, random forest, static graph neural networks, relational graph neural networks, dynamic graph neural networks, heterogeneous transformers, and the proposed framework. The proposed framework achieves the strongest minority-class F1 in numerical comparison, improves temporal stability in later snapshots, and reduces false-negative risk under severe class imbalance. The results show that heterogeneous feature alignment, temporal propagation, and pseudo-risk calibration are complementary rather than interchangeable. The study contributes to business and data analytics by linking graph-based artificial intelligence with operational fraud governance, threshold policy, and risk-based resource allocation. It also offers practical guidance for exchanges, payment processors, compliance teams, and digital-asset risk managers that need to convert large-scale blockchain data into timely and defensible fraud intelligence.
