International Journal of Science, Engineering and TechnologyAuthors: Vijay Saini, Jitender Kumar
Abstract: Digital payment ecosystems have expanded both the volume and complexity of transaction data, widening the opportunity for fraud while raising customer, operational, and regulatory expectations that automated decisions remain explainable. This review consolidates research on credit card fraud detection spanning rule-based systems, classical statistical learning, ensemble and gradient-boosted methods, deep learning architectures, and explainable artificial intelligence (XAI), and synthesises these strands into FraudDetectNet, a layered pipeline in which data preprocessing, feature reduction, classification, explanation generation, and monitoring are treated as interdependent rather than separable functions. Particular attention is given to Shapley Additive Explanations (SHAP), used here not only as a post-hoc diagnostic but as an upstream feature-selection mechanism that compresses high-dimensional transaction data while preserving discriminative power. The review also examines evaluation practice for severely imbalanced, temporally ordered fraud data, arguing for precision-recall and cost-sensitive measures over plain accuracy. Outcomes reported in the primary reference study underlying this review — feature-space reduction from 380 to 120 variables, 97.6% accuracy, 99.0% fraud-class recall, and training-time reduction from 185.3 to 54.1 seconds — are presented as evidence of feasibility rather than generalised guarantees, and extensions toward graph-based, federated, and continual learning are outlined.
DOI: http://doi.org/10.5281/zenodo.20813647
