Toward Auditable Fraud Detection: Combining Graph Features, Model Explanations, and Agentic Case Investigation