Scholarlink

Organizations face major challenges in detecting employee expense fraud because of scarce data, intricate fraud patterns, and the requirement for explainable results. This paper implements a novel application that integrates Case-Based Reasoning (CBR) with Large Language Models (LLMs) to address these challenges. Our system represents expense activities as spatiotemporal events, uses LLMs to generate fraud detection rules within a constrained function space, and applies CBR to retrieve similar cases, minimize hallucinations, and improve explainability. The system implements a complete CBR cycle—retrieve similar fraud patterns, reuse detection rules, revise rules through LLM interaction, and retain verified cases. We evaluated the system with more than 200,000 real-world expense events and the results show that the integration of CBR with LLMs effectively constrains hallucinations while generating high-quality, explainable fraud detection rules. This approach offers a practical solution for applying AI in high-stakes domains requiring reliability and explainability.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating Case-Based Reasoning with LLM for Expense Fraud Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xiaoyu Ge,
  • Jiao Xu

摘要

Organizations face major challenges in detecting employee expense fraud because of scarce data, intricate fraud patterns, and the requirement for explainable results. This paper implements a novel application that integrates Case-Based Reasoning (CBR) with Large Language Models (LLMs) to address these challenges. Our system represents expense activities as spatiotemporal events, uses LLMs to generate fraud detection rules within a constrained function space, and applies CBR to retrieve similar cases, minimize hallucinations, and improve explainability. The system implements a complete CBR cycle—retrieve similar fraud patterns, reuse detection rules, revise rules through LLM interaction, and retain verified cases. We evaluated the system with more than 200,000 real-world expense events and the results show that the integration of CBR with LLMs effectively constrains hallucinations while generating high-quality, explainable fraud detection rules. This approach offers a practical solution for applying AI in high-stakes domains requiring reliability and explainability.