|国家科技期刊平台
首页|期刊导航|南京大学学报(自然科学版)|基于多阶近邻约束的深度不完整多视图聚类方法

基于多阶近邻约束的深度不完整多视图聚类方法OA北大核心CSTPCD

Deep incomplete multi-view clustering based on multi-order neighborhood constraint

中文摘要英文摘要

多视图聚类是重要的无监督学习方法之一,然而在实际应用中很难获取完整的多视图数据,导致不完整多视图聚类问题.大多数已有的不完整多视图聚类方法只考虑了视图的属性信息,而忽视了数据结构信息对聚类的影响,使提取的特征不能充分表示原始数据的潜在结构.针对以上问题,提出一种基于多阶近邻约束的深度不完整多视图聚类方法.首先,利用具有自注意力机制的深度自编码器获取带有视图间信息交互的深层次隐含特征,并采用加权融合的方式获取视图的公共语义信息;然后,对于不完整多视图中的缺失数据,利用多视图的公共表示进行补全;最后,提出一种多阶近邻约束机制,该机制考虑不完整多视图数据的深层结构信息,利用多视图的互补性构建近似完整的近邻图,引导编码器学习更紧致、更有判别性的高级语义特征.在公共数据集上的实验结果证明了所提方法的有效性.

Multi-view clustering is an important unsupervised learning method.However,in real applications,it is difficult to obtain complete multi-view data,which leads to incomplete multi-view clustering problem.Most of the existing incomplete multi-view clustering methods only consider the attribute information of views,but ignore the influence of structure information on clustering,resulting in extracted features cannot fully represent the latent structure of the original data.To address these problems,in this paper,a deep method based on multi-order neighborhood constraints is proposed for incomplete multi-view clustering.Firstly,the deep autoencoder with self-attention is used to obtain the rich complex latent features with cross-view information interaction,and the weighted fusion approach is employed to learn the consistency common information of views.Then,in incomplete multi-view settings,the missing data are fixed up by the consistency common representation of multi-views data.Finally,the multi-order neighborhood constraint mechanism is proposed,which considers the deep structural information within incomplete views and constructs an approximate complete neighborhood graph using the complementarity of multi-views,guiding the encoder to learn more compact and discriminative high-level semantic features.Experimental results show that the proposed method is effective.

王梅;王伟东;刘勇;于源泽

东北石油大学计算机与信息技术学院,大庆,163318中国人民大学高瓴人工智能学院,北京,100049

计算机与自动化

不完整多视图聚类自注意力结构信息多阶近邻

incomplete multi-view clusteringself-attentionstructure informationmulti-order neighborhood

《南京大学学报(自然科学版)》 2024 (001)

53-64 / 12

国家自然科学基金(51774090,62076234),黑龙江省博士后科研启动金资助项目(LBH-Q20080)

10.13232/j.cnki.jnju.2024.01.006

评论