EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

REAL-TIME FINANCIAL FRAUD DETECTION USING A FEATURE-LIGHT TEMPORAL DYNAMIC GRAPH NEURAL NETWORK WITH CONTRASTIVE LEARNING

YiWen Liang, YaNan Jiao

Journal of Trends in Finance and Economics · 01.01.2024 · 1 (1) · 76-84

DOI: 10.61784/jtfe4076 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Trends in Finance and Economics, online erschienen am 01.01.2024. Verfasst von YiWen Liang und 1 weiteren Autor:innen. Die Arbeit wurde bisher 12-mal zitiert (Crossref). Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Financial fraud detection is a rare-event, non-stationary graph classification problem in which relational patterns can change faster than labels become available. This paper presents FT-TDGCL, a feature-light temporal directed graph neural network with contrastive learning for near-real-time anti-money-laundering screening. Unlike approaches that depend on the 166 anonymized Elliptic attributes, the proposed pipeline uses only the public directed transaction edge list and labels. It derives 18 auditable topology features, applies incoming and outgoing graph filters at one and two hops, and fuses a node embedding with a causal gated recurrent snapshot context. During training, two independently channel-masked and noise-perturbed graph-filtered views are aligned with an NT-Xent objective in addition to class-balanced binary cross-entropy. We evaluate on the real Elliptic Bitcoin graph with 203,769 transactions, 234,355 payment flows, and 49 temporal steps. A strict chronological split uses steps 1–29 for training, 30–34 for validation and threshold selection, and 35–49 for testing; unknown labels participate only in topology construction. Across three seeds, FT-TDGCL obtains a test PR-AUC of 0.2152±0.0935 versus 0.1655±0.0333 without contrastive learning, a 30.0% relative gain. It reaches recall 0.5429±0.0444, while its F1 of 0.2344±0.0393 does not surpass every fixed-seed baseline, exposing a calibration trade-off rather than a universal improvement. Cached-feature inference over 16,670 labeled test nodes requires 6.95±0.64 ms on CPU. All metrics, predictions, hashes, and code are generated by the accompanying reproducibility package; no synthetic observations or hand-entered experimental results are used.

Zitieren

APA

Liang, Y., & Jiao, Y. (2024). REAL-TIME FINANCIAL FRAUD DETECTION USING A FEATURE-LIGHT TEMPORAL DYNAMIC GRAPH NEURAL NETWORK WITH CONTRASTIVE LEARNING. Journal of Trends in Finance and Economics, 1(1), 76-84. https://doi.org/10.61784/jtfe4076

Vancouver

Liang Y, Jiao Y. REAL-TIME FINANCIAL FRAUD DETECTION USING A FEATURE-LIGHT TEMPORAL DYNAMIC GRAPH NEURAL NETWORK WITH CONTRASTIVE LEARNING. Journal of Trends in Finance and Economics. 2024;1(1):76-84. doi:10.61784/jtfe4076

BibTeX
@article{Liang2024,
  author = {Liang, YiWen and Jiao, YaNan},
  title = {REAL-TIME FINANCIAL FRAUD DETECTION USING A FEATURE-LIGHT TEMPORAL DYNAMIC GRAPH NEURAL NETWORK WITH CONTRASTIVE LEARNING},
  journal = {Journal of Trends in Finance and Economics},
  year = {2024},
  volume = {1},
  number = {1},
  pages = {76--84},
  doi = {10.61784/jtfe4076}
}
Kostenlos starten →