TEMPLAR fraud a verifier grounded calibrated and cost sensitive framework for transaction fraud detection

Abstract

Transaction fraud detection remains difficult because fraudulent events are highly imbalanced, temporally drifting, cost asymmetric, and often supported by limited auditability in model decisions. Existing tabular, tree-based, and deep fraud detectors can achieve strong predictive performance, but they often provide weak verifier-grounded reasoning, incomplete probability calibration, and limited integration with cost-sensitive decisioning. This paper proposes TEMPLAR-Fraud, a verifier-grounded, calibrated, and cost-sensitive framework for transaction fraud detection. The framework combines SAINT-based transaction representation learning, CatBoost triage, routed symbolic template reasoning, deterministic rationale verification, Bayesian stacking, one-dimensional optimal transport calibration, and validation-selected cost-sensitive thresholding. Evaluation was conducted on BAF Base and IEEE-CIS Fraud Detection using chronological train/validation/test splits, internal testing, same-dataset temporal future-slice evaluation, calibration analysis, cost-sensitive utility assessment, robustness testing, adversarial stress testing, computational-efficiency measurement, and explanation-safety analysis. On the internal chronological test sets, TEMPLAR-Fraud achieved AUROC/AUC-PR/F1 scores of 0.918/0.498/0.557 on BAF Base and 0.972/0.701/0.747 on IEEE-CIS Fraud Detection. Under same-dataset temporal future-slice evaluation, it retained AUROC/AUC-PR/F1 scores of 0.901/0.452/0.518 and 0.951/0.642/0.687, while reducing ECE after calibration to 0.014 and 0.013 and producing the highest reported net savings under the stated experimental cost model. These findings suggest that framework-level integration of routed prediction, verified rationale generation, calibration, and cost-sensitive thresholding can improve fraud-detection evaluation under controlled benchmark protocols.