摘要
现有大多数对比图聚类算法存在以下问题:生成节点表示时忽略了浅层网络提取的底层特征和底层结构信息;未充分利用高阶邻居节点信息;未结合置信度信息与拓扑结构信息来构建正样本对。为解决以上问题,提出了基于多层特征融合与增强的对比图聚类算法。该算法首先融合不同层次网络提取的节点特征,以补充节点的底层结构信息;其次,通过节点间的局部拓扑相关性和全局语义相似度聚合节点信息,以增强节点表示的上下文约束一致性;最后,联合置信度信息和拓扑结构信息构建更多高质量正样本对,提高簇内表示一致性。实验结果表明,CGCMFFE在四种广泛使用的聚类评价指标上表现出优异的性能。理论分析和实验研究验证了CGCMFFE中节点底层特征、高阶邻居节点信息、置信度和拓扑结构信息的关键作用,证明了CGCMFFE的优越性。
The majority of existing contrastive graph clustering algorithms face the following issues:they ignore the low-level features and structural information extracted by shallow networks when generating node representation.The algorithms neither fully utilize high-order neighbor node information nor integrate confidence information with topological structure information to construct positive sample pairs.To address the above issues,this paper proposed a contrastive graph clustering algorithm based on multi-level feature fusion and enhancement.The algorithm firstly integrated node features extracted from different network layers to enrich the low-level structural information of nodes.It then aggregated node information through the local topological correlations and global semantic similarities between nodes to enhance the contextual constraint consistency of node representations.Finally,combining confidence information and topological structure information,the algorithm constructed more high-quality positive sample pairs to improve the consistency of intra-cluster representation.The experimental results show that CGCMFFE has excellent performance on four widely used clustering evaluation metrics.Theoretical analysis and experimental study underscore the crucial role of low-level node features,highorder neighbor node information,confidence,and topological structure information in the CGCMFFE algorithm,providing evidence for its superiority.
作者
李志明
魏贺萍
张广康
尤殿龙
Li Zhiming;Wei Heping;Zhang Guangkang;You Dianlong(School of Information Science&Engineering,Yanshan University,Qinhuangdao Hebei 066004,China;The Key Laboratory for Software Engineering of Hebei Province,Yanshan University,Qinhuangdao Hebei 066004,China;The Key Laboratory for Computer Virtual Technology&System Integration of Hebei Province,Yanshan University,Qinhuangdao Hebei 066004,China;Shenzhen Research Institute of Yanshan University,Yanshan University,Shenzhen Guangdong 518063,China)
出处
《计算机应用研究》
北大核心
2025年第6期1749-1754,共6页
Application Research of Computers
基金
国家自然科学基金面上项目(62276226)
河北中央引导地方项目(236Z7725G)
河北省重点研发计划项目(20375001D)。
关键词
多层特征融合
对比图聚类
无监督学习
multi-level feature fusion
contrastive graph clustering
unsupervised learning
