Graph-Based Financial Fraud Detection: A Property-Centric Survey of Static, Temporal, Cycle-Aware, and Generative Methods and Their Evaluation

Financial fraud is increasingly coordinated across multiple accounts and transactions, meaning that its evidence often lies in relationships between accounts, the sequence of events, and, in schemes such as money laundering and wash trading, closed loops of transfers rather than in individual records. This review examines graph-based methods for detecting such fraud, covering static and temporal graph neural networks, graph anomaly detection, cycle- and motif-aware methods, generative models, and the evaluation practices used to assess them. The review draws on 147 sources identified through a documented literature search protocol. Four findings emerge. First, no single method family performs best under all conditions: tree-based models remain highly competitive on feature-rich datasets, whereas graph-based and temporal models offer greater advantages when discriminative signals reside in relational structures or transaction sequences. Second, detecting cyclic laundering is fundamentally a representational problem: standard message-passing architectures have limited ability to capture cycle structures, while recent performance gains have been associated with representations that preserve the directed multigraph structure of transactions rather than simply increasing model complexity. Third, evaluation practices, including temporal leakage and threshold selection using test data, can inflate reported performance and undermine the reliability of comparisons. Fourth, a taxonomy of eight method families evaluated against four properties required for coordinated, temporal, and partially observed fraud detection—temporal awareness, probabilistic generation, cycle awareness, and intrinsic explainability—reveals that none of the reviewed families provides all four properties. This identifies a central research gap in the current literature. Building on these findings, the review proposes cyclic fraud as a structural stress test for future detection methods and provides an evaluation checklist to address the methodological weaknesses identified.

Authors

Institutions

Publication Details

Journal
Journal of Computing Theories and Applications
Published
2026-11-17
DOI
https://doi.org/10.62411/jcta.18143
Primary Topic
Advanced Graph Neural Networks
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Graph-Based Financial Fraud Detection: A Property-Centric Survey of Static, Temporal, Cycle-Aware, and Generative Methods and Their Evaluation

Daniel T. Ale, E. O. Ogunti, A. A. Ponnle, K. B. Adedeji et al.
6 links
Journal of Computing Theories and Applications
Advanced Graph Neural Networks
article

Graph-Based Financial Fraud Detection: A Property-Centric Survey of Static, Temporal, Cycle-Aware, and Generative Methods and Their Evaluation

Daniel T. Ale, E. O. Ogunti, A. A. Ponnle, K. B. Adedeji, S. A. Oyetunji
article en

Abstract

Financial fraud is increasingly coordinated across multiple accounts and transactions, meaning that its evidence often lies in relationships between accounts, the sequence of events, and, in schemes such as money laundering and wash trading, closed loops of transfers rather than in individual records. This review examines graph-based methods for detecting such fraud, covering static and temporal graph neural networks, graph anomaly detection, cycle- and motif-aware methods, generative models, and the evaluation practices used to assess them. The review draws on 147 sources identified through a documented literature search protocol. Four findings emerge. First, no single method family performs best under all conditions: tree-based models remain highly competitive on feature-rich datasets, whereas graph-based and temporal models offer greater advantages when discriminative signals reside in relational structures or transaction sequences. Second, detecting cyclic laundering is fundamentally a representational problem: standard message-passing architectures have limited ability to capture cycle structures, while recent performance gains have been associated with representations that preserve the directed multigraph structure of transactions rather than simply increasing model complexity. Third, evaluation practices, including temporal leakage and threshold selection using test data, can inflate reported performance and undermine the reliability of comparisons. Fourth, a taxonomy of eight method families evaluated against four properties required for coordinated, temporal, and partially observed fraud detection—temporal awareness, probabilistic generation, cycle awareness, and intrinsic explainability—reveals that none of the reviewed families provides all four properties. This identifies a central research gap in the current literature. Building on these findings, the review proposes cyclic fraud as a structural stress test for future detection methods and provides an evaluation checklist to address the methodological weaknesses identified.

Journal of Computing Theories and ApplicationsVol. 4(2)
Federal University of Technology (NG)
Advanced Graph Neural Networks
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon