Advanced Predictive Modeling and Real-Time Anomaly Detection for Unemployment Insurance Fraud Mitigation: A Multi-Model Machine Learning Framework for Public Benefit Systems
International Journal of Computer Applications Technology and Research · 2025
The unprecedented surge in Unemployment Insurance (UI) claims, particularly following the COVID-19 pandemic, has exposed critical vulnerabilities in public benefit systems, leading to staggering financial losses attributable to fraudulent activities.Traditional fraud detection methods, predominantly reliant on static, rule-based systems and postpayment audits, are ill-equipped to counter the sophisticated, large-scale, and adaptive nature of modern fraud schemes.This paper introduces the Predictive Anomaly and Network Detection for Operational Risk Abatement (PANDORA) framework, a novel, multi-modal machine learning architecture designed for real-time fraud mitigation in UI systems.PANDORA integrates three specialized analytical modules: (1) a supervised learning component utilizing an XGBoost classifier trained on historical fraud data to generate claim-level propensity scores; (2) an unsupervised anomaly detection component employing an Isolation Forest algorithm to identify novel and emergent fraud typologies not present in historical data; and (3) a graph neural network (GNN) module for uncovering complex, collusive fraud rings through network analysis of claimant, employer, and infrastructural data.These modules operate in concert, feeding into an ensemble meta-learner that calculates a unified Composite Risk Score (CRS) for each claim.This score facilitates a dynamic, risk-based triage system, enabling real-time decision-making: auto-approval, manual review, or immediate denial.We present a simulated implementation using a large-scale synthetic dataset modeled on real-world claim characteristics, demonstrating that PANDORA achieves a 28% improvement in F1-score and a 42% reduction in false positive rates compared to traditional benchmarks.The framework's design addresses critical considerations including model interpretability through SHAP (SHapley Additive exPlanations), scalability, and a continuous learning feedback loop, presenting a robust and adaptive solution to a pressing public administration challenge.