Year: 2026
Type: article
Abstract: Introduction: The rapid growth of digital financial transactions has increased exposure to fraud and highlighted the limitations of traditional decision support systems (DSS). Rule-based and scorecard-driven approaches struggle to operate effectively under conditions of extreme class imbalance, evolving fraud patterns, and regulatory requirements for transparency and auditability. This study aimed to design and evaluate an AI-driven DSS for credit card fraud detection that integrated machine learning, class-imbalance mitigation, explainability, and cost-sensitive decision logic to support operational fraud analysis. Method: The study used the publicly available Kaggle Credit Card Fraud dataset comprising 284,807 transactions, of which 0.17% were fraudulent. A stratified 80/20 split yielded a hold-out test set of 4,000 transactions, including 7 fraud cases. Class imbalance was addressed using the Synthetic Minority Oversampling Technique (SMOTE) applied exclusively within training folds. Logistic Regression, Random Forest, and XGBoost models were trained and evaluated. Given the extreme imbalance, Precision–Recall Area Under the Curve (PR–AUC) was treated as the primary evaluation metric, with recall, precision, F1-score, and ROC–AUC reported as complementary measures. Performance uncertainty was quantified using bootstrap confidence intervals. Probability calibration was applied to XGBoost outputs, and a cost-sensitive threshold optimisation framework was used to define operational risk tiers. SHapley Additive exPlanations (SHAP) were incorporated to support model transparency. Results: XGBoost achieved the highest PR–AUC (0.568) compared with Random Forest (0.512) and Logistic Regression (0.341), while exhibiting comparable recall (0.571) to Random Forest on the imbalanced test set. Bootstrap confidence intervals indicated substantial uncertainty due to the limited number of fraud cases, underscoring the need for cautious interpretation of point estimates. Calibration improved probability reliability, and cost-sensitive thresholding produced consistent low-, medium-, and high-risk classifications. SHAP analysis provided global and local explanations of model behaviour, although the use of anonymised PCA-derived features constrained interpretability. Conclusion: The findings demonstrated that integrating SMOTE, XGBoost, calibration, and SHAP within an AI-driven DSS supported improved fraud prioritisation under severe class imbalance while maintaining transparency and operational feasibility. Rather than replacing existing systems, the proposed framework illustrated how AI models can augment DSS by enabling calibrated risk scoring, uncertainty-aware evaluation, and analyst-oriented triage in regulated financial environments.
Authors: Yazeed Alsuhaibany, Shabir Ahmad
Institution: Al Yamamah University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7212931663