Network Analytics for Anti-Money Laundering -- A Systematic Literature Review and Experimental Evaluation

TL;DR

This work addresses the fragmented literature on network analytics for anti-money laundering by providing a systematic review of 97 papers and a standardized experimental framework. It demonstrates that combining network topology with node features yields predictive gains, while graph neural networks face stability and imbalance challenges in AML settings. The study benchmarks manual features, shallow embeddings, and deep representations on the Elliptic and IBM-AML datasets, revealing limited but meaningful improvements from advanced methods and cautionary results from synthetic data. The authors also release open-source code to promote standardized evaluation and replication, aiming to accelerate robust adoption of network analytics in AML practice.

Abstract

Money laundering presents a pervasive challenge, burdening society by financing illegal activities. The use of network information is increasingly being explored to effectively combat money laundering, given it involves connected parties. This led to a surge in research on network analytics for anti-money laundering (AML). The literature is, however, fragmented and a comprehensive overview of existing work is missing. This results in limited understanding of the methods to apply and their comparative detection power. This paper presents an extensive and unique literature review, based on 97 papers from Web of Science and Scopus, resulting in a taxonomy following a recently proposed fraud analytics framework. We conclude that most research relies on expert-based rules and manual features, while deep learning methods have been gaining traction. This paper also presents a comprehensive framework to evaluate and compare the performance of prominent methods in a standardized setup. We compare manual feature engineering, random walk-based, and deep learning methods on two publicly available data sets. We conclude that (1) network analytics increases the predictive power, but caution is needed when applying GNNs in the face of class imbalance and network topology, and that (2) care should be taken with synthetic data as this can give overly optimistic results. The open-source implementation facilitates researchers and practitioners to extend this work on proprietary data, promoting a standardised approach for the analysis and evaluation of network analytics for AML.

Figures (18)

• Figure 1: The search and filtering method with the number of remaining papers between brackets.
• Figure 2: The number of papers incorporated in the full query and after filtering per year.
• Figure 3: The evolution of the model building blocks over the years. Since there are many modelling methods, the graphs are split.
• Figure 4: The evolution of the evaluation metrics over the years. Since there are many metrics, the graphs are split.
• Figure 5: For the feature engineering and shallow representation methods, the obtained network features/embedding is added to the intrinsic ones.