Deployment-Oriented Interpretable Fraud Detection via Hybrid Explainable Boosting Machines with Concept–Raw Fusion on the IEEE-CIS Benchmark

Fraud detection models often achieve a strong ranking performance through black-box ensembles, but operational deployment also requires calibration, low explanation cost, and auditable scoring logic. This study develops an interpretable fraud-detection pipeline for IEEE-CIS by combining a 63-variable causal concept bank with teacher-guided additive Explainable Boosting Machine (EBM) students. The concept bank summarizes the temporal state, entity history, novelty/reuse, identity missingness, and aggregate deviation. Experiments use a chronological out-of-time split and a stricter pseudo-entity-disjoint holdout. In the main three-seed evaluation, the CatBoost predictive ceiling and XGBoost teacher achieved PR-AUC 0.489 ± 0.001 and 0.478 ± 0.003, respectively. Among interpretable models, concept-only EBM reached 0.189 ± 0.000, raw-only EBMs reached 0.372 ± 0.005 (top-k = 8) and 0.383 ± 0.002 (top-k = 12), and hybrid EBMs reached 0.407 ± 0.003 (top-k = 8) and 0.407 ± 0.004 (top-k = 12), consistently improving over matched raw-only additive baselines. The final top-k = 8 hybrid reduced input features from 154 to 71, achieved about 9.7× faster inference than XGBoost, remained close to XGBoost in ECE-15 calibration (0.01587 vs. 0.01611) while having a higher Brier score, and produced native local explanations far faster than XGBoost + SHAP. The results position CatBoost as the predictive ceiling and hybrid EBM as a benchmark-supported, deployment-relevant interpretable compromise for applied financial risk-screening workflows, rather than as a production-validated fraud-monitoring system.