基于混合阶相似性的多视图聚类:一个广义的视角  

Multiview Clustering by Hybrid-Order Affinity:A Generalized Perspective

作　　者：陈曼笙 任骊安 王昌栋[1,2] 黄栋 赖剑煌[1] CHEN Man-Sheng;REN Li-An;WANG Chang-Dong;HUANG Dong;LAI Jian-Huang(School of Computer Science and Engineering,Sun Yat-sen University,Guangzhou 510006;Guangdong Provincial Key Laboratory of Intellectual Property and Big Data,Guangzhou 510006;College of Mathematics and Informatics,South China Agricultural University,Guangzhou 510642)

机构地区：[1]中山大学计算机学院,广州510006 [2]广东省知识产权大数据重点实验室,广州1510006 [3]华南农业大学数学与信息学院,广州510642

出　　处：《计算机学报》2024年第7期1453-1468,共16页Chinese Journal of Computers

基　　金：中国国家重点研发计划(2021YFF1201202);国家自然科学基金(62276277,61976097);广东省知识产权大数据重点实验室(2018B030322016)资助.

摘　　要：多视图聚类已经被广泛研究,它能够采用可用的多源信息来实现更好的聚类性能.然而,大多数之前的工作仍存在两个不足:(1)它们通常关注多视图属性特征的场景,很少留意到多视图属性图数据;(2)它们主要尝试发现一致的结构或多个视图之间的关系,而忽略了多视图观测之间潜在的高阶相关性。为了解决这些问题,我们从广义角度出发,提出了一种新颖的方法,称为混合阶相似性的多视图聚类(Multiview Clustering by Hybridorder Affinity,MCHA).它将结构图和多视图属性特征巧妙融合,同时考虑了低秩概率相似性图和混合阶的相关性.具体而言,我们通过图过滤策略构建了一组保留几何结构的视图特定的平滑表示.同时,我们将从平滑表示中学习得到的多视图概率相似性图堆叠成一个张量,并对该张量给予低秩属性的约束.这可以很好地恢复视图间更高阶的相关性.在八个基准数据集上的实验表明,我们所提出的MCHA方法具有最先进的有效性.Multiview clustering capable of adopting the available multisource information is extensively studied to achieve better clustering performance.However,most previous literatures still suffer from two limitations.(1)They often concentrate on the scenario of multiview attributes,paying little attention to the multiview attributed graph data.(2)They mainly attempt to discover a consensus structure or the relationships between multiple views,mostly neglecting the underlying higher-order correlations between multiview observations.To tackle these problems,we propose a novel method called Multiview Clustering by Hybrid-order Affinity(MCHA)from a generalized perspective,where the structural graph and the multiview attributes are seamlessly fused,and the low-rank probability affinity graphs with hybrid-order correlations are simultaneously considered.Specifically,a set of view-specific smooth representations preserving the geometrical structure is constructed by means of the graph filtering strategy.Meanwhile,we stack multiview probability affinity graphs learned from smooth representations into a tensor constrained by low-rank property,so that the higher-order correlations can be well recovered.Experiments on eight benchmark datasets indicate the state-of-the-art effectiveness of the proposed MCHA method.

关 键 词：多视图聚类 概率相似性图 低秩张量 高阶相关性 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]