基于加权GraphSAGE和生成对抗网络的医保欺诈识别方法

Medical fraud detection method based on weighted GraphSAGE and generative adversarial network

医保欺诈行为分析与识别是医疗保险基金监管中最为重要的工作,对保障医保基金安全与可持续发展有着非常重要的意义.为保证医保欺诈行为识别的准确度,需充分挖掘医保数据中的患者信息.然而,对于缺乏欺诈样本的不平衡医保数据集,目前常用的医保欺诈识别模型的泛化能力不佳且性能下降.因此,本文提出了一种基于加权GraphSAGE和生成对抗网络的医保欺诈识别方法.该方法融合了患者就诊关系特征表示与基于加权GraphSAGE算法的患者特征提取,并结合生成对抗网络构建识别模型.实验证明,本方法大大提升了模型的识别性能.同时,我们将所提方法与元路径向量、图卷积神经网络、图注意力网络、多层图注意力网络和超图自适应聚类网络等先进主流识别模型对比发现,本文提出的识别方法在召回率、精确率、F1值和准确率等指标下表现也更好;在不同数据规模和不同正负样本比例下,模型性能稳定,有较好的泛化性.

Medicare fraud analysis and detection is the most critical task in medical fund su-pervision,essential to ensure medical funds’security and sustainable development.To ensure the accuracy of medicare fraud detection,one needs to explore the patient information in the data fully.However,many detection models have poor generalization ability and degraded per-formance when dealing with medicare imbalanced datasets that lack fraud samples.Therefore,this paper proposes a medicare fraud detection method based on weighted GraphSAGE and gen-erative adversarial network.This method combines the representation of relationship features of patient visits with weighted GraphSAGE algorithm-based patient feature extraction and employs generative adversarial network to construct detection models.Experiments demonstrate that the proposed method significantly improves the recognition performance of the model.Meanwhile,we compare the proposed method with advanced mainstream recognition techniques such as meta-path vectors,convolutional neural network,graph attention network,heterogeneous graph attention network and one-class adversarial nets.The proposed recognition method performs better in Recall,Precision,F1-score and Accuracy.Moreover,its performance remains stable under different data sizes and various positive and negative sample ratios,offering better gener-alization.