基于多关系异构谱图神经网络的财务报表欺诈检测

Financial Statement Fraud Detection Based on Multi-Relational Heterogeneous Spectral Graph Neural Networks

摘要: 财务报表欺诈严重破坏资本市场的健康运行与资源配置，构建精准高效的欺诈检测模型具有重要的现实意义。随着图神经网络(GNN)的发展，基于企业关联网络的异常检测成为重要研究方向。然而，现有GNN模型多基于同质性假设(即相连节点具有相似特征或标签)，难以有效应对欺诈者为掩盖造假行为而刻意构建的大量异质性伪装连接。同时，传统方法往往忽略了企业间存在的供销、投资等多维复杂的网络关系。针对上述挑战，本文提出了一种基于多关系异构谱图神经网络特征提取的检测框架(M-RHGDF)。该框架首先通过边缘感知模块预测节点间的同质或异质关联，进而将原始的多关系图动态分裂为特定的正负子图。随后，引入基于谱图理论的可调Beta小波图神经网络(BWGNN)，针对不同子图进行特定频段的特征提取，有效分离代表正常模式的低频信号与揭示异常行为的高频信号。最后，通过聚合不同关系网络下的频域表征，实现对欺诈节点的精准分类。本文在经过数据预处理、字段脱敏且无时间依赖性的真实企业关联数据集FDCompCN上进行了详尽评估。实验结果表明，M-RHGDF模型在AUC、GMean、F1-macro及召回率等核心指标上均显著优于现有主流基线模型，充分证明了其在解决异构图欺诈检测问题中的优越性与鲁棒性。

Abstract: Financial statement fraud poses a severe threat to the healthy operation and resource allocation of capital markets, making the development of accurate and efficient fraud detection models highly crucial. With the advancement of Graph Neural Networks (GNNs), anomaly detection based on corporate association networks has emerged as a promising research direction. However, most existing GNN models rely on the homophily assumption—which presumes that connected nodes share similar characteristics—and thus struggle to handle the extensive heterophilous connections deliberately forged by fraudsters for camouflage. Furthermore, traditional methods often overlook the multiplex, complex relations among enterprises, such as supply, distribution, and investment links. To tackle these challenges, this paper proposes a Multi-Relation Heterogeneous Graph Detection Framework (M-RHGDF) based on spectral graph feature extraction. Specifically, an edge-aware module is first utilized to predict whether the relations between nodes are homophilous or heterophilous, dynamically splitting the original multi-relation graph into specific positive and negative subgraphs. Subsequently, an adjustable Beta Wavelet Graph Neural Network (BWGNN) grounded in spectral graph theory is introduced to perform frequency-specific feature extraction on different subgraphs, effectively separating low-frequency signals that represent normal patterns from high-frequency signals indicative of anomalous behaviors. Finally, the spectral representations across various relational networks are aggregated to achieve a precise classification of fraudulent nodes. Extensive evaluations are conducted on FDCompCN, a real-world corporate association dataset characterized by rigorous pre-processing, anonymized fields, and temporal independence. Experimental results demonstrate that M-RHGDF significantly outperforms existing mainstream baseline models across core metrics including AUC, GMean, F1-macro, and Recall, thoroughly validating its superiority and robustness in addressing heterophilous graph fraud detection.