2025, Vol. 12, No. s3. - go to content...

Permanent address of this page - https://resources.today/en/02faor325.html

Метаданные этой статьи так же доступны на русском языке

DOI: 10.15862/02FAOR325 (https://doi.org/10.15862/02FAOR325)

Full article in PDF format (file size: 472.6 KB)

For citation:

Rusakov K.A. Synthesis of graph neural networks and natural language processing methods to detect anomalous patterns in P2P transfer chains as a risk-based supervision tool. Russian journal of resources, conservation and recycling. 2025; 12(s3). Available at: https://resources.today/PDF/02FAOR325.pdf (in Russian). DOI: 10.15862/02FAOR325

Synthesis of graph neural networks and natural language processing methods to detect anomalous patterns in P2P transfer chains as a risk-based supervision tool

Rusakov Kirill Alekseevich
Financial University under the Government of the Russian Federation, Moscow, Russia
E-mail: 245308@edu.fa.ru

Abstract. The rapid growth of interpersonal electronic transfers amid the proliferation of the Faster Payments System and digital financial platforms is creating a fundamentally new financial fraud risk landscape. Traditional anti-fraud monitoring systems based on signature rules and threshold filters demonstrate diminishing effectiveness as fraudulent schemes adapt to automated controls. This study analyzes the potential of synthesizing graph neural networks and natural language processing methods to detect anomalous patterns in P2P transfer chains in the context of risk-based supervision. This paper systematizes key types of fraudulent schemes in the P2P transfer sphere, analyzes the architectural features of graph neural networks applicable to transaction graph analysis, and explores the potential for integrating NLP components to enhance behavioral analytics. A conceptual framework for applying a hybrid GNN-NLP approach, depending on the anomaly type and data source, has been developed. The methodological basis consists of a systems approach and a comparative analysis of algorithms, enabling the comparison of neural network architectures based on accuracy, response speed, and suitability for processing financial data in real time. The information base utilizes analytical reports from FinCERT, data from Kaspersky Lab and Positive Technologies on the financial sector threat landscape for 2024–2025, as well as scientific publications in the field of graph and linguistic transaction analysis. It has been established that graph representation of transaction data allows for the identification of structural patterns inaccessible through element-by-element analysis, while NLP components enable the recognition of semantic markers of manipulative influence in the text fields of transactions.

Keywords: graph neural networks; natural language processing; peer-to-peer (P2P) transfers; financial fraud; anti-fraud; risk-based oversight; Faster Payments System; behavioral analytics; machine learning; transaction graph

Download article in PDF format