全文下载:
202505018.pdf
文章编号: 1672-6987(2025)05-0142-10 DOI: 10.16351/j.1672-6987.2025.05.018
朱淼, 陈卓*, 杜军威, 胡强(青岛科技大学 信息科学技术学院, 山东 青岛 266061)
摘要: 目前属性图异常检测算法通常使用自编码器对图进行重构,难以充分利用图中丰富的信息进行图表示学习,且此类算法使用重构损失作为目标判定函数,不面向异常信息优化节点表示,因而限制了检测性能。针对以上问题,本工作提出一种基于联合自监督学习的属性图异常检测算法。该算法利用构建的全局、子图两种异常检测机制,挖掘图中节点在属性和结构信息中蕴含的异常特征。此外,算法通过构建联合损失函数,将节点异常判定信息融入异常节点表示,提升捕获异常信息的能力。实验结果表明,该算法对属性图异常节点的检测效果在5个公开数据集上比其他5种基线方法均有显著提升。
关键词: 异常检测; 自监督学习; 图神经网络; 图表示学习; 联合训练
中图分类号: TP 183 文献标志码: A
引用格式: 朱淼, 陈卓, 杜军威, 等. 基于联合自监督学习的属性图异常检测算法[J]. 青岛科技大学学报(自然科学版), 2025, 46(5): 142-151.
ZHU Miao, CHEN Zhuo, DU Junwei, et al. Anomaly detection on attribute networks based on joint self-supervised learning[J]. Journal of Qingdao University of Science and Technology(Natural Science Edition), 2025, 46(5): 142-151.
Anomaly Detection on Attribute Networks Based on Joint Self-Supervised Learning
ZHU Miao, CHEN Zhuo, DU Junwei, HU Qiang(College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China)
Abstract: At present, the graph anomaly detection models usually use the autoencoder to reconstruct the graph, which is difficult to make full use of the rich information in the graph to learn the graph representation. And the reconstruction loss is used as the target determination function of the algorithm, they do not optimize the node representation, which limits the performance of the algorithm. To solve the above problems, this paper proposes a graph anomaly detection algorithm based on joint self-supervised learning, which deeply explores the differences between graph anomaly nodes and global or adjacent subgraph nodes in attribute and structure information, and constructs two anomaly detection mechanisms of global and subgraph, making full use of the local, global and interaction information between graph nodes. By constructing a joint loss function, the model integrates the node anomaly determination information into the graph abnormal node representation learning, and improves the ability of the model to capture abnormal information. The experimental results show that the detection effect of the proposed algorithm on the abnormal nodes of the graph is significantly improved on the five public data sets compared with the other five baseline methods.
Key words: anomaly detection; self-supervised learning; graph neural network; graph representation learning; joint training
收稿日期: 2024-11-26
基金项目: 国家自然科学基金项目 (62172249).
作者简介: 朱淼 (1998—), 女, 硕士研究生. * 通信联系人.