Embedded Files
What Was Achieved, and What Comes Next
Finding 1: Random Forest achieved F1=0.8586, ROC-AUC=0.9794 — correctly identifying 82 of 98 fraud cases with only 11 false alarms.
Finding 1: Random Forest achieved F1=0.8586, ROC-AUC=0.9794 — correctly identifying 82 of 98 fraud cases with only 11 false alarms.
Finding 2: SHAP identified V14, V12, and V4 as the three dominant fraud predictors. Low V14 values serve as the single strongest fraud signal.
Finding 2: SHAP identified V14, V12, and V4 as the three dominant fraud predictors. Low V14 values serve as the single strongest fraud signal.
Finding 3: LIME produced a fully audit-ready explanation at P(Fraud)=1.0000 — the first instance-level compliance record satisfying EU AI Act Article 13.
Finding 3: LIME produced a fully audit-ready explanation at P(Fraud)=1.0000 — the first instance-level compliance record satisfying EU AI Act Article 13.
The Core Takeaway
The Core Takeaway
This research demonstrates, empirically, that the long-assumed trade-off between fraud detection accuracy and regulatory transparency is an architectural problem — not a fundamental one. When compliance requirements are embedded into the pipeline from the start rather than bolted on afterward, both goals are simultaneously achievable. The framework produced ROC-AUC of 0.9838 (Stacking Ensemble) and full EU AI Act-compliant documentation within the same pipeline.
Acknowledged Limitations
Acknowledged Limitations
Dataset Temporal Scope → Two-day window from 2013; modern fraud patterns (digital wallets, crypto-linked cards) not represented.
PCA Anonymisation → V1–V28 cannot be mapped to business concepts like merchant category or geography by compliance officers.
Static Resampling → SMOTE applied at a single snapshot; no drift detection for shifting fraud distributions.
Performance Gap → F1=0.8586 is 6.26% below the Iqbal et al. baseline (F1=0.9159) due to fixed default hyperparameters.
LIME Instability → Random perturbation sampling produces marginally different outputs per call; reproducibility requires stabilisation.
Five Directions for Future Research
Five Directions for Future Research
Real-time streaming compliance generation for high-frequency transaction systems
Real-time streaming compliance generation for high-frequency transaction systems
Behavioral trust metrics to measure whether SHAP/LIME improves stakeholder comprehension
Behavioral trust metrics to measure whether SHAP/LIME improves stakeholder comprehension
Adversarial transaction patterns in training to resist evasion by adaptive fraudsters
Adversarial transaction patterns in training to resist evasion by adaptive fraudsters
Systematic hyperparameter optimisation to close the F1 gap relative to published baselines
Systematic hyperparameter optimisation to close the F1 gap relative to published baselines
Cross-dataset validation beyond Kaggle for generalisability across institutions and regions
Cross-dataset validation beyond Kaggle for generalisability across institutions and regions
Page updated
Report abuse