arxivcs.LGcs.AI2026-07-21
Toward Auditable Fraud Detection: Combining Graph Features, Model Explanations, and Agentic Case Investigation
Fraud detection systems must scale with rising transaction volume while remaining explainable and reviewable. We study a layered pipeline on the PaySim dataset that combines a gradient-boosted classifier, graph-derived structural features, an autoencoder-based anomaly signal, Tre…