Adaptive Graph Neural Analytics for Cryptocurrency Anomaly Detection under Limited Labeled Data
Main article
Abstract
Cryptocurrency anomaly detection has become a central problem in AI analytics because illicit transactions, exchange abuse, laundering chains, ransomware payments, and coordinated mixing behavior evolve faster than manually labeled investigation data. Existing graph neural methods have improved blockchain risk screening, yet many remain dependent on balanced labels, static graph assumptions, or a single node type. This paper develops an Adaptive Graph Neural Analytics framework for cryptocurrency anomaly detection under limited labeled data. The proposed framework constructs dynamic heterogeneous transaction graphs from address, transaction, entity, and temporal interaction evidence; learns risk-sensitive node representations through relation-aware graph encoding; and updates decision boundaries through confidence-guided pseudo-labeling, temporal consistency regularization, and investigator-feedback calibration. Instead of treating scarce labels as a secondary inconvenience, the framework places label scarcity at the center of the model design. A benchmark-style evaluation based on Elliptic++-inspired public transaction structures and controlled label-ratio scenarios shows that the proposed approach improves anomaly-class F1 by 8.4 to 15.9 percentage points over temporal GCN, graph autoencoder, and class-weight multilayer perceptron baselines when only 1% to 10% of labels are available. Ablation analysis further indicates that heterogeneous alignment, adaptive temporal sampling, and pseudo-label verification contribute complementary performance gains. The paper contributes analytically transparent architecture for AI-assisted blockchain compliance, a data-efficient learning strategy for highly imbalanced cryptocurrency graphs, and an operational discussion of deployment risks including drift, adversarial adaptation, investigator workload, and auditability.
