Self-Adaptive Graph Analytics for Real-Time Fraud Detection in Transaction Networks

Financial fraud in large scale transaction networks has
become more sophisticated, and traditional static and rule-based
systems are not anymore adequate to deal with dynamic and
evolving fraudulent behaviors. The issue is that conventional
graph-based and machine learning models struggle with the
changing topology in real-time financial data streams which must
be adaptable to fast changes in topology and concept drift. This
study proposes a Self-Adaptive Graph Analytics (SAGA)
framework that is able to dynamically model transaction
networks with temporal graph learning, reinforcement-driven
topology adaptation and continuous online feedback. The goal is
to increase accuracy in detecting fraud in near real-time and
S.Thiruselvi
Assistant Professor
PERI College of Arts and Science
Chennai, Tamil Nadu, India
thiruselviprakash23@gmail.com
S. Arunarani
Assistant Professor
Department of Computer
Applications,Faculty of Science and
Humanities, SRM Institute of Science and
Technology. Chennai, Tamil Nadu, India
arunaras@srmist.edu.in
synthetic transaction networks[2]. Traditional methods of
fraud detection (based on static rules, thresholds or historical
features) are facing issues in dynamic patterns and tend to
result in late fraud detection with high false positive rate[3].
As a result, intelligent systems that are adaptive and scalable
and that can learn from continuously changing streams of data
and detect subtle, hidden patterns of fraud in real-time manner
are in an urgent need.
keep the inference time low. The methodology uses the temporal
graph neural networks(TGNN), meta-learning based self
adaptation and graph based anomaly scoring mechanism to
evolve with the dynamics of transactions continuously.
Experimental evaluation using Elliptic Bitcoin Transaction
Dataset shows that SAGA achieves an F1 score of 0.90 and an
AUC score of 0.96 and outperforms 5 state-of-the-art temporal
and static graph baselines with an inference latency of less than
12 ms per transaction. The outcomes prove SAGA has the ability
to effectively capture evolving fraud patterns, reduce false
alarms, and enable real time operational deployment. This study
concludes by stating that the adaptive graph analytics is a
promising way towards scalable, transparent and resilient fraud
detection in modern financial systems.
Keywords: Graph Neural Networks; Temporal Graph Learning;
Fraud Detection; Real-Time Analytics; Adaptive Learning;
Transaction Networks; Finan