基于交易行为表征学习的企业财务欺诈检测方法

A Corporate Financial Fraud Detection Method Based on Transaction Behavior Representation Learning

公司间关联交易已成为实施财务欺诈的常用手段。传统定量分析方法将每家公司视为一个独立个体,未考虑交易各方间复杂关系。为此,提出了基于混合交易行为特征和集成机器学习模型的企业财务欺诈检测方法。首先,构建知识图谱,从企业交易数据中提取关联交易特征,并与常用财务特征整合在一起。其次,提出结合决策树(DT)、随机森林(RF)和Adaboost算法的财务欺诈检测集成框架。取混合特征作为输入,子模型对每笔交易是否欺诈进行投票,通过硬投票或软投票聚合方法达成最终决策。在上市公司的现实交易数据集上的实验表明,混合特征能够增强财务欺诈检测性能,且所提框架通过集成多样化模型和不同的投票机制,在欺诈交易检测中实现了92.46%的受试者工作特征曲线下面积(AUC),检测性能显著优于单个分类器,有助于促进企业可持续增长,协助监管机构维护市场秩序。

Inter-company related transactions have become a common way for executing financial fraud.Traditional quantitative analysis methods treat each company as an independent entity,failing to explore the complex relationships a-mong the transaction parties.To address this,we propose a corporate financial fraud detection method based on mixed transaction behavior features and an ensemble machine learning model.First,a knowledge graph is constructed to extract related transaction features from corporate transaction data,which are then integrated with common financial features.Subsequently,we introduce an ensemble framework for financial fraud detection that combines Decision Tree(DT),Tandom Forest(RF),and Adaboost algorithms.The hybrid features serve as input,and the sub-models vote on whether each transaction is fraudulent,reaching a final decision through hard or soft voting aggregation methods.Experi-ments on real transaction datasets of listed companies demonstrate that the hybrid features enhance financial fraud detec-tion performance.The proposed framework achieves a 92.46%AUC in detecting fraudulent transactions by integrating diverse models and various voting mechanisms,significantly outperforming individual classifiers.This method facilitates sustainable corporate growth and assists regulatory agencies in maintaining market order.