Temporal Graph Learning for Robust Detection of Coordinated Fraud in Dynamic Transaction Networks
Coordinated fraud can produce repeated, synchronized transfers whose signatures are obscured by heterogeneous activity, missing edges, and evolving behavior. This paper studies a lightweight temporal graph learning strategy based on account-level transaction attributes, static neighborhood descriptors, and eight-bin temporal coordination statistics. Static and temporal feature variants are each trained with the same simulated edge-deletion augmentation to disentangle the effects of timing descriptors and robustness training. Experiments are conducted on a controlled, fully specified synthetic benchmark containing 4,800 accounts, approximately 15,000 transfers per run, and simulator-assigned coordinated actors; the benchmark is not a real-world labeled fraud dataset. Three independent runs use chronological, nonoverlapping train/validation/test snapshots. The edge-augmented static graph achieves test average precision 0.631, while the identically augmented temporal feature model reaches 0.602; their nonaugmented counterparts score 0.571 and 0.569, respectively. At 20% and 40% test-edge removal, augmented static-graph average precision is 0.518 and 0.389, respectively. A matched augmentation control, fixed-budget top-100 precision, and timestamp shuffling separate the roles of training robustness and within-snapshot timing. The study provides an efficient, transparent protocol for evaluating coordinated-pattern detection in evolving transaction networks.
Authors
- Sofia Lindqvist
- Maya Svensson
- Johan Berg
Institutions
- Uppsala University (SE)
- Stockholm University (SE)
Publication Details
- Journal
- International Journal of Advanced Engineering and Technology Research
- Published
- 2026-11-21
- DOI
- https://doi.org/10.54097/7nnpma38
- Primary Topic
- Advanced Graph Neural Networks
- Type
- article
- Field-Weighted Citation Impact
- 0.00