基于自编码器及对比损失的图聚类方法  

Graph Clustering Based on Auto-Encoder and Contrastive Loss

作　　者：王静红[1,2,3] 王慧 袁绰[4] Wang Jinghong;Wang Hui;Yuan Chuo(College of Computer and Cyber Security,Hebei Normal University,Shijiazhuang 050024,China;Hebei Provincial Key Laboratory of Network and Information Security,Shijiazhuang 050024,China;Hebei Provincial Engineering Research Center for Supply Chain Big Data Analytics and Data Security,Shijiazhuang 050024,China;Hebei Institute of Engineering Technology,Shijiazhuang 050020,China)

机构地区：[1]河北师范大学计算机与网络空间安全学院,河北石家庄050024 [2]河北省网络与信息安全重点实验室,河北石家庄050024 [3]供应链大数据分析与数据安全河北省工程研究中心,河北石家庄050024 [4]河北工程技术学院,河北石家庄050020

出　　处：《南京师大学报(自然科学版)》2025年第1期75-84,共10页Journal of Nanjing Normal University(Natural Science Edition)

基　　金：河北省自然科学基金项目(F2021205014、F2019205303);河北省高等学校科学技术研究项目(ZD2022139);中央引导地方科技发展资金项目(226Z1808G);河北师范大学科技类研究基金项目(L2023J05);河北师范大学研究生创新资助项目(XCXZZSS202315)。

摘　　要：图聚类根据图数据的内在关系找到组或社区,是数据分析中一项重要的研究问题.近年来,基于自编码器的方法能够获得有效节点属性表示,但未融合结构信息.由于图神经网络的广泛应用,基于半监督图卷积网络和图自编码器的模型能够融合结构信息,与传统聚类方法相比取得了较好的效果,但标记数据和卷积操作代价昂贵.因此,本文提出了一种基于自编码器及对比损失的图聚类方法.首先该方法使用简单的多层感知器设计自编码器,预训练自编码器学习节点属性表示.其次结合影响对比损失学习图嵌入表示,融合丰富的图结构信息,然后同时迭代优化嵌入表示和自监督聚类任务.最后,使用多个引文网络数据集与基准模型进行对比实验.实验表明,聚类性能得到有效提升,并且参数敏感性分析和变体实验验证了影响对比损失和自监督聚类的有效性.Graph clustering,which can find communities or groups based on the intrinsic relationships of graph data,is an important research problem in data analysis.In recent years,Auto-Encoder based methods have effectively extracted node attribute representations,but do not include structural information.Due to the widespread application of graph neural networks,the fusion of structural information based on semi-supervised graph convolutional networks and graph Auto-Encoder models has achieved better results compared to traditional clustering methods.However,labeling data and using convolution operations are expensive.This paper proposed a graph clustering method based on Auto-Encoder and Contrastive Loss(GC-AECL).Firstly,the model used a simple multilayer perceptron to design the Autoencoders,and pretrained the Auto-Encoder learning node attribute representation.Then the model combined the influence contrastive loss to enrich structural information to learn the graph embedded representation.And then the model iteratively optimized the embedded representation and self-supervised clustering tasks at the same time.Finally,the experiments were compared with the benchmark models on multiple citation network datasets.The experimental results showed that the clustering performance had been improved,and parameter sensitivity analysis and variants study had been conducted to verify the effectiveness of impact contrastive loss and self-supervised clustering.

关 键 词：图聚类 自编码器 影响对比损失 图嵌入 自监督聚类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]