%0 Journal Article
%T 基于多关系不平衡图的图神经网络欺诈检测<br>Graph Neural Network Fraud Detection Based on Multi-Relation Imbalanced Graph
%A 赵佳玉
%J Software Engineering and Applications
%P 752-762
%@ 2325-2278
%D 2023
%I Hans Publishing
%R 10.12677/SEA.2023.126073
%X 近年来，为了挖掘图结构数据中包含的丰富的关系信息，基于图神经网络的欺诈检测方法引起了人们的广泛关注。然而，对于节点标签分布严重偏斜并且存在多种关系的图数据，传统的图神经网络可能会表现不佳。为了解决多关系不平衡图上的欺诈检测问题，本文提出了一种基于关系感知的图神经网络模型(RA-GNN)。首先，对目标节点进行基于节点相似度的Top-p邻居节点采样以改善节点标签分布不平衡问题；其次，在每种关系内部，使用节点级注意力机制，加强关键邻居节点对中心节点嵌入的影响，聚合邻居信息得到每种关系下的邻居表示；最后，使用边级注意力机制自适应地学习每种关系下邻居表示的重要性，实现关系感知的邻居信息聚合，得到中心节点的嵌入向量并应用于欺诈检测任务。在真实欺诈检测数据集YelpChi和Amazon上的实验结果表明，本文提出的RA-GNN模型与其他基于图神经网络的基线模型相比具有良好的欺诈检测性能，并且所使用的核心模块有显著的增强效果。<br />
In recent years, fraud detection methods based on graph neural networks have attracted much attention due to the rich relational information contained in graph-structured data. However, for graph data with severely skewed node label distribution and multiple relationships, traditional graph neural networks may perform poorly. In order to solve the fraud detection problem on multi-relationship unbalanced graphs, this paper proposes a relationship-aware graph neural network-based model (RA-GNN). First, Top-p neighbor node sampling based on node similarity is performed on target nodes to improve the problem of imbalanced node label distribution; second, within each relationship, a node-level attention mechanism is used to enhance the influence of key neighbor nodes on the central node embedding, and aggregate neighbor information to obtain neighbor representation under each relationship; Finally, an edge-level attention mechanism is used to adaptively learn the importance of the neighbor representation under each relationship to achieve relationship-awareness. Finally, using the edge-level attention mechanism to adaptively learn the importance of the neighbor representation under each relationship, the relationship-aware neighbor information is aggregated to obtain the embedding vector of the central node and applied to the fraud detection task. Experiments on both benchmark and real-world graph-based fraud detection tasks demonstrate that the proposed RA-GNN apparently outperforms other baselines.
%K 欺诈检测，多关系不平衡图，图神经网络，注意力机制<br>Fraud Detection
%K Multi- Relation Imbalanced Graph
%K Graph Neural Network
%K Attention Mechanism
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=77129