From Structure to Suspicion: Cross-Network Detection of Fraudulent Behavior in Real-World Signed Graphs

This study presents a supervised anomaly detection (AD) approach for real-world transaction networks, emphasizing cross-network generalization and structural feature extraction. Two signed and weighted Bitcoin trading platforms, namely Bitcoin Alpha and Bitcoin OTC, are analyzed by modeling user interactions as directed graphs, where edge weights encode trust and distrust ratings. Fraudulent behavior is detected by leveraging a combination of five graph-based centrality metrics alongside topological information derived from clustering. These features are used to train machine learning classifiers, tasked with identifying suspicious nodes. We conducted cross-network experiments in which the model is trained on one dataset and tested on the other, thereby simulating real-world generalization scenarios without incorporating temporal data. A comprehensive evaluation reveals that the proposed approach outperforms prior methods reported in the literature. Specifically, this study demonstrates high detection capability, low false alarm rates, and strong overall robustness across both datasets.