Detection, Attribution, Narration: An End-to-End Pipeline for Explainable Money Mule Identification

2026-07-20

AI Summary

This study addresses the challenges of strong data heterogeneity and insufficient interpretability in identifying money mule accounts in financial fraud detection. The authors propose an end-to-end identification framework that integrates a LightGBM classifier—trained on 280-dimensional multi-source features—with TreeSHAP for feature attribution and a large language model to generate natural language explanations. This work presents the first synergistic application of TreeSHAP and large language models in a regulated financial context. The approach significantly enhances interpretability while maintaining high coverage, achieving a detection rate increase from 61% to 89% after deployment. Monthly valid alerts rose from 211 to 302, and an additional 60% of previously undetected risky accounts were identified, substantially reducing analysts’ cognitive load.

Abstract

Money mule accounts are critical facilitators of financial fraud, yet detecting them at scale remains challenging due to the heterogeneous nature of transactional and behavioural data. We present an end-to-end pipeline for customer-level mule detection comprising three stages: (1) a LightGBM classifier trained on 280 engineered features spanning transaction patterns, account demographics, network topology, and temporal behaviour; (2) a TreeSHAP attribution layer that decomposes each prediction into feature contributions; and (3) a large language model (LLM) module that converts SHAP attributions into analyst-facing natural-language narratives. We evaluate across three open-weight LLM families and assess explanation quality through analyst feedback. In a live production deployment, the system achieves a yield rate of 89%, up from 61% under the incumbent rule-based system, with monthly alert volume expanding from 211 to 302, reflecting broader true-positive coverage rather than increased noise. This corresponds to a 60% incremental adverse detection beyond existing review workflows, substantially outperforming the rule-based approach. Qualitative feedback from analysts indicates that LLM-generated narratives reduce cognitive load during alert triage. We further discuss implications of deploying LLM-augmented explainability in regulated financial environments.

Problem

• money mule detection
• financial fraud
• explainable AI
• transactional data
• behavioral heterogeneity

Innovation

• Explainable AI
• Money Mule Detection
• TreeSHAP Attribution
• Large Language Model
• End-to-End Pipeline