The pseudonymous nature of blockchain transactions poses a significant challenge for identifying fraudulent activity in decentralized financial systems. This study presents a comprehensive framework for classifying
Bitcoin wallets as fraudulent or legitimate by integrating multisource data, graph-based transaction modeling, and machine learning. Our methodology builds upon publicly available datasets—namely Elliptic++, Chainabuse, and a curated sample of recent transactions—and integrates structural, temporal, and monetary features extracted from the Bitcoin transaction graph. Through systematic experiments across three distinct labeling scenarios, we demonstrate that ensemble methods such as Random Forests offer strong performance even under label
noise, achieving F1-scores up to 0.92. Moreover, an explainability framework grounded, in SHAP values, is used to systematically analyze feature contributions and elucidate behavioral patterns associated with financial
fraud. Our approach bridges empirical robustness with forensic insight, contributing a scalable, transparent toolset for blockchain compliance and risk analysis.
