"Amazon"

"Graph Neural Networks (GNNs) have become the predominant approach for graph fraud detection due to their intrinsic capability to handle graph-structured data and effectively capture complex relational patterns in fraudulent behaviors. However, existing GNN-based graph fraud detection models face limitations: homophily-based models struggle with handling heterogeneous relationships in fraud graphs, while heterophily-based models typically model only a single attribute- or structural-space, leading to constrained detection performance. To address these issues, this paper introduces DualH-FDNet, a semi-supervised graph fraud detection model based on dual-space heterogeneous relation analysis. This model represents user relationships as multi-relational heterogeneous directed graphs and employs a multi-layer graph convolutional architecture. Each convolutional layer consists of three modules: (1) Heterogeneity Learning Module: Utilize the label information of labeled nodes in relational subgraphs to learn heterogeneity separately in the attribute-space and structural-space, and achieve feature interaction of dual-space heterogeneity through a weighted fusion strategy. (2) Cross-Space Graph Aggregation Module: It computes attention weights based on fused heterophily representations and updates node representations via multi-relational graph aggregation. (3) Prototype-Guided Classification Module: It constructs category prototypes using labeled node representations and labels, guiding the classification of unlabeled nodes through prototype learning. Additionally, to tackle the challenges of scarce labeled data and label imbalance, the model utilizes balanced sampling strategies for semi-supervised training. Experimental results show that on the YelpChi and Amazon datasets, DualH-FDNet improves Recall by 0.9626% and 0.6444%, respectively, and AUC by 0.8594% and 0.1479% compared to the best-performing baseline models among nine comparative models. This study offers a novel solution for fraud detection in complex heterogeneous graph environments. The code and data are available at https:\/\/github.com\/AyomF\/DualH-FDNet."

图神经网络（Graph Neural Networks，GNNs）已成为图欺诈检测的主流方法，因其具备处理图结构数据的固有能力，可有效捕捉欺诈行为中的复杂关联模式。然而，现有的基于GNN的图欺诈检测模型存在诸多局限：基于同质性（homophily）的模型难以处理欺诈图中的异质性关联，而基于异质性（heterophily）的模型通常仅对单一属性空间或结构空间进行建模，导致检测性能受限。为解决上述问题，本文提出了DualH-FDNet，一种基于双空间异质性关联分析的半监督图欺诈检测模型。该模型将用户关联表征为多关系异质性有向图，并采用多层图卷积架构。每个卷积层包含三个模块：（1）异质性学习模块（Heterogeneity Learning Module）：利用关系子图中带标签节点的标签信息，分别在属性空间与结构空间中独立学习异质性特征，并通过加权融合策略实现双空间异质性的特征交互。（2）跨空间图聚合模块（Cross-Space Graph Aggregation Module）：基于融合后的异质性表征计算注意力权重，并通过多关系图聚合更新节点表征。（3）原型引导分类模块（Prototype-Guided Classification Module）：利用带标签节点的表征与标签构建类别原型，通过原型学习引导未标记节点的分类任务。此外，为应对带标签数据稀缺与标签不平衡的挑战，该模型采用平衡采样策略开展半监督训练。实验结果表明，在YelpChi与Amazon数据集上，相较于9个对比模型中的最优基线模型，DualH-FDNet的召回率（Recall）分别提升0.9626%与0.6444%，曲线下面积（Area Under Curve，AUC）分别提升0.8594%与0.1479%。本研究为复杂异质性图环境下的欺诈检测提供了全新解决方案。相关代码与数据集可于https://github.com/AyomF/DualH-FDNet获取。

搜集汇总

背景与挑战

背景概述

Amazon数据集是一个用于图神经网络欺诈检测研究的标准数据集，包含用户对乐器的评论数据，用户被标记为欺诈或良性，并定义了三种基于产品、评分和文本相似性的关系类型。该数据集共有11,944个节点，欺诈节点占比为6.87%，适用于评估半监督图欺诈检测模型的性能。