TRANSACTION GRAPH ANALYSIS UNDER TEMPORAL NONSTATIONARITY

Read full text Summarise with VIRAT
Authors: Danylo Yefimov
Cited by: 0
References: 9

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The study involved an experimental evaluation of illicit-transaction detection on the Elliptic Data Set, a temporally annotated directed graph of Bitcoin transactions.Multiple model families were examined, including Logistic Regression, Random Forest, Graph Convolutional Network (GCN), GraphSAGE (Graph Sample and Aggregate), and Chebyshev Spectral Graph Convolutional Network (ChebNet), along with weighted ensemble combinations.Hyperparameters were tuned exclusively on the validation period, and decision thresholds were calibrated prior to final evaluation on the future test data.The transductive semi-supervised setting allowed access to the full graph structure while restricting labels to the training subset.Adversarial validation was conducted to assess distributional differences between temporal splits and revealed near-perfect separability between the pre-test and test data, indicating a covariate shift.Statistical testing further identified a structural break aligned with a known dark-market shutdown event, accompanied by a significant change in the proportion of illicit transactions.Across all model families, strong validation performance consistently degraded under the temporally shifted test distribution, and ensemble strategies did not eliminate this generalization gap.The findings demonstrate that temporal nonstationarity constitutes a dominant challenge in transaction-graph learning.Access to relational structure and unlabeled nodes, while beneficial for representation learning, does not inherently ensure robustness to covariate shift.The results highlight the necessity of shift-aware validation, diagnostic testing, and adaptation strategies for reliable deployment of graph-based anti-money laundering systems.

Fields of Study & Subjects

Advanced Graph Neural Networks Privacy-Preserving Technologies in Data Cybercrime and Law Enforcement Studies

Publication details

Journal / source
Наука і техніка сьогодні
Issue
2(56)
Publication date
21 March 2026
Language
EN