An AI-Enabled Anti-Fraud Risk Early-Warning System and Empirical Evidence: A Big-Data and Multi-Model Framework
Main article
Abstract
Telecom fraud has evolved into an industrialized and highly organized form of financial crime, exposing limitations in rule-based monitoring and siloed model deployment within commercial banks. This study proposes a lifecycle-oriented risk early-warning system that integrates three complementary modules: (i) a pre-event account-opening risk rating model driven by multi-source big-data profiling, (ii) an in-process transaction-level detector built on LightGBM for real-time identification of anomalous patterns, and (iii) a post-event linkage mining module that combines LPA-based community detection with knowledge-graph association analysis to uncover organized fraud networks and expand investigative leads. Using production data from Bank A, we demonstrate that the integrated framework improves both coverage and operational actionability, strengthens closed-loop risk governance across pre-event prevention, mid-event monitoring, and post-event tracing, and delivers measurable reductions in reported fraud risk relative to peer benchmarks. The proposed architecture offers a deployable blueprint for financial institutions seeking scalable, data-driven, and continuously updatable anti-fraud risk management.
