TY  - EJOU
AU  - Raj, Rahul 

TI  - FRAUD-LENS: Hybrid Deep Learning for Real-Time Unemployment Insurance Fraud Detection via Temporal Behavioral Drift
T2  - Journal on Artificial Intelligence

PY  - 2026
VL  - 8
IS  - 1
SN  - 2579-003X

AB  -  <b>Background:</b> Unemployment Insurance (UI) fraud represents one of the most costly threats to social benefit integrity, with the U.S. DOL/ETA estimating improper payments exceeding $45 billion between 2020 and 2023. Existing detection systems fail to model the temporal evolution of claiming behavior or the relational topology connecting fraudulent actors across employer-claimant networks. This study introduces FRAUD-LENS, a hybrid deep learning framework delivering interpretable and scalable fraud detection for large-scale federal UI systems. <b>Methods:</b> FRAUD-LENS integrates three coordinated architectural modules: a Bidirectional Long Short-Term Memory (BiLSTM) network encoding temporal claim behavior sequences, a Graph Attention Network (GAT) modeling employer-claimant relational topology to surface fraud ring membership, and a calibrated XGBoost ensemble on 47 engineered tabular features. A novel feature engineering paradigm, Temporal Behavioral Drift (TBD), is introduced with formal convergence guarantees to quantify deviation from a claimant’s personal behavioral baseline across seven behavioral dimensions. Predictions from all three streams are fused through a fully specified cross-modal attention mechanism. A Fraud Probability Cascade (FPC) score provides continuous triage aligned with DOL/ETA protocols, with thresholds calibrated through an empirical three-step process. <b>Results:</b> Evaluated on a de-identified dataset of 1.2 million UI claim sequences from 18 states (2018–2023), with GAN-based synthetic augmentation applied exclusively to the training partition, FRAUD-LENS achieves F1-score of 95.0% (95% CI: [94.4%, 95.6%]), AUC-ROC of 0.981, precision of 95.3%, and recall of 94.8%, outperforming all evaluated baselines including a Temporal Transformer (F1: 90.8%). Temporal validation (train 2018–2021, test 2022–2023) yields F1 of 94.3%; cross-state held-out validation across five unseen states yields mean F1 of 93.4% (range: 90.5%–96.4%). Fairness analysis reveals consistent performance across age groups and geographic strata. The cold-start subgroup (fewer than four certification weeks) achieves F1 of 84.1% and is identified as the primary performance gap. <b>Conclusions:</b> FRAUD-LENS demonstrates strong and generalizable performance for UI fraud detection. The TBD methodology and FPC scoring mechanism show promise for extension to banking, healthcare, and other means-tested benefit program fraud contexts, though further external validation is warranted before broad deployment.
KW  - Unemployment insurance fraud detection; hybrid deep learning; graph attention network; bidirectional LSTM; temporal behavioral drift; XGBoost ensemble; fraud probability cascade; social benefit integrity; anomaly detection; DOL/ETA

DO  - 10.32604/jai.2026.083202