A Unified Explainable Machine Learning Framework for Fraud Detection Across Payment and Cryptocurrency Transaction Systems

Abstract

Financial fraud detection remains a critical challenge due to extreme class imbalance and heterogeneous transaction patterns across payment and cryptocurrency systems. In this study, we present a unified and explainable machine learning framework for binary fraud detection across three benchmark datasets: PaySim mobile money transactions, Elliptic Bitcoin transactions, and a real-world credit card fraud dataset. Logistic Regression, Random Forest, and XGBoost classifiers are implemented under a consistent preprocessing and evaluation framework. Model performance is assessed using ROC–AUC, precision, recall, F1-score, and accuracy metrics, while SHAP-based feature attribution provides interpretability of model predictions. Experimental results demonstrate that ensemble-based models, particularly Random Forest and XGBoost, consistently outperform Logistic Regression, achieving high discriminatory power under severe class imbalance. The findings provide empirical benchmarks for fraud detection across heterogeneous transactional systems and highlight the practical value of explainable ensemble models for financial risk assessment.