Neuro-Symbolic Artificial Intelligence for Secure, Explainable, and Intelligent Banking and FinTech Systems: Fraud Detection, Credit Risk, Financial Forecasting, and Transaction Security (Scholars Journal of Engineering and Technology)

Summary

• On the IEEE-CIS fraud benchmark the proposed framework achieved an F1 score of 0.79, with precision of 0.81, recall of 0.77 and a false positive rate of 0.9 percent.p.19
• The disparate impact ratio for the proposed framework was 0.91, compared with 0.84 for gradient boosting and 0.81 for the deep learning baseline.p.20
• The framework achieved a mean absolute error of 0.87 percent on the one-day-ahead forecast and 2.14 percent on the five-day-ahead forecast.p.21
• Detection rates were 98.2 percent for card testing, 95.7 percent for account takeover and 95.3 percent for rapid transfer fraud.p.22
• Rule satisfaction rate was 99.6 percent for transaction security and 99.2 percent for credit.p.24
• The framework reduced the false positive rate from 1.3 percent to 0.9 percent, and analysts preferred its explanations in 78 percent of cases.p.25

Key insights

1. The framework combines a symbolic rule engine with neural learning in a single pipeline serving fraud detection, credit risk, forecasting and transaction security.p.4
2. Reported literature results include DeepProbLog reaching F1 = 0.67 versus XGBoost 0.61 on fraud detection.p.6
3. The gradient boosting baseline reached an F1 score of 0.71 and the deep learning baseline without symbolic reasoning reached 0.74 on fraud detection.p.19
4. Directional accuracy was 61.3 percent for one-day-ahead and 57.8 percent for five-day-ahead forecasts.p.21
5. A rule-only baseline achieved 100 percent rule satisfaction but a much lower AUC of 0.69.p.21
6. Explanation faithfulness was 0.89 for the proposed framework.p.22
7. The average increase in mean absolute error between calm and volatile windows was 0.31 percentage points.p.22