管理评论

管理评论 ›› 2025, Vol. 37 ›› Issue (10): 197-209.

• 会计与财务管理 • 上一篇    

基于知识图谱的上市公司财务欺诈识别方法

杨玥敏1, 王超2, 陈浩智3, 张卫国4   

  1. 1. 华南理工大学机械与汽车工程学院, 广州 510640;
    2. 广东工业大学经济学院, 广州 510520;
    3. 华南理工大学工商管理学院, 广州 510641;
    4. 深圳大学管理学院, 深圳 518057
  • 收稿日期:2023-12-25 发布日期:2025-11-18
  • 作者简介:杨玥敏（通讯作者），华南理工大学机械与汽车工程学院辅导员，硕士；王超，广东工业大学经济学院讲师，博士；陈浩智，华南理工大学工商管理学院博士研究生；张卫国，深圳大学管理学院教授，博士生导师，博士。
  • 基金资助:
    科技创新2030-“新一代人工智能”重大项目（2020AAA0108404）；广东省自然科学基金面上项目（2023A1515012494）。

A Method Based on Knowledge Graph to Detect Listed Companies’ Fraud in Financial Statement

Yang Yuemin1, Wang Chao2, Chen Haozhi3, Zhang Weiguo4   

  1. 1. School of Mechanical & Automotive Engineering, South China University of Technology, Guangzhou 510640;
    2. School of Economics, Guangdong University of Technology, Guangzhou 510520;
    3. School of Business Administration, South China University of Technology, Guangzhou 510641;
    4. School of Management, Shenzhen University, Shenzhen 518057
  • Received:2023-12-25 Published:2025-11-18

PDF (PC)

18

摘要： 上市公司的财务欺诈行为对金融市场稳定造成了重大威胁，已有上市公司财务欺诈识别模型的相关研究很少考虑上市公司之间及其与金融实体的关联特征，这些关系是否可以提高欺诈识别的准确性有待探索。本研究提出考虑上市公司关联关系的财务欺诈识别方法，该方法利用知识图谱刻画上市公司之间的共同人员、一致行动、共同关联方关系，通过知识推理提取同构网络，将同构网络特征与公司自身特征相结合，使用轻量级梯度提升机算法（LightGBM）构建欺诈识别模型。基于我国A股上市公司2016—2020年的真实数据进行实证检验，结果表明，关联关系网络特征有助于提高财务欺诈识别的准确性。与其他机器学习方法相比，结合网络特征的LightGBM取得了更好的性能。

关键词: 财务欺诈, 知识图谱, 机器学习, 网络中心性

Abstract: Financial statement fraud of listed companies is seriously threatening the stability of the global economy. Existing studies rarely consider the relationship between listed companies and other financial entities. It remains to be explored whether these relationships can enhance the accuracy of fraud detection. This paper proposes a detection method considering the relationship between listed companies. This method constructs a knowledge graph to depict common management, concerted action, and common related party relationships. Through knowledge reasoning, a homogeneous network is extracted. Integrating the features from the homogeneous network with the company’s own characteristics, a fraud detection model is constructed using the Light Gradient Boosting Machine (LightGBM). This paper conducts an empirical test based on the real data of China’s A-share listed companies from 2016 to 2020. The empirical results indicate that the network features related to interrelationships contribute to improved accuracy in financial statement fraud detection. In comparison with other machine learning methods, LightGBM methods considering the relationships exhibits better performance.

Key words: financial statement fraud, knowledge graph, machine learning, network centrality