张量学习诱导的多视图谱聚类  

Tensor Learning Induced Multi-View Spectral Clustering

作　　者：陈曼笙 蔡晓莎 林家祺 王昌栋[1,3] 黄栋 赖剑煌[1] CHEN Man-Sheng;CAI Xiao-Sha;LIN Jia-Qi;WANG Chang-Dong;HUANG Dong;LAI Jian-Huang(School of Computer Science and Engineering,Sun Yat-sen University,Guangzhou 510006;School of Mathematics(Zhuhai),Sun Yat-sen University,Zhuhai,Guangdong 519000;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]中山大学数学学院(珠海),广东珠海519000 [3]广东省知识产权大数据重点实验室,广州510006 [4]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1期52-68,共17页Chinese Journal of Computers

基　　金：国家自然科学基金(62276277);广东省自然科学基金(2022B1515120059);广东省知识产权大数据重点实验室(2018B030322016)资助。

摘　　要：现有的方法将通过张量奇异值分解(t-SVD)正则化的低秩表示应用到多视图子空间聚类中,取得了令人印象深刻的聚类性能.然而,它们都具有以下两个共同的缺点:(1)他们专注于探索样本之间的关系以构建表征,然后将其堆叠为张量,其计算复杂度至少为O(n2logn);(2)他们总是直接在整合的表征上运行标准的谱聚类算法,而忽略了不同表征对最终聚类结果的先验知识.为了解决这些问题,本文提出了一种新颖的张量学习诱导的多视图谱聚类(TLIMSC)方法,其中同时探索了空间聚类结构和互补信息.具体来说,该方法将关联样本和簇关系的多视图谱嵌入表示堆叠成张量,计算复杂度最终变为O(n logn).然后,将学习到的带有不同自适应置信度的表征与最终的一致聚类结果联系起来.在五个数据集上的广泛实验证明了TLIMSC所具有的有效性和高效性.Low-rank representation coefficients regularized by the tensor-Singular Value Decom-position(t-SVD)scheme for multi-view subspace clustering have achieved impressive performances.However,all of them suffer from the following two common demerits.(1)They focus on exploring the relationships among samples to construct representations which are then stacked to be a tensor,whose computational complexity is at least O(n²logn);(2)They always deploy the standard spectral clustering algorithm directly on the integrated representation,neglecting the prior knowledge of different representations towards the final results.To tackle these problems,we propose a novel Tensor Learning Induced Multi-view Spectral Clustering(TLIMSC)approach,where the spatial cluster structures and complementary information are simultaneously explored.Specifically,multi-view spectral embedding representations related from samples to clusters are focused to be stacked in a tensor,where the complexity finally becomes O(nlogn).Later,a bridge would be built to connect the learned representations carrying different adaptive confidences with the final consensus results.Extensive experiments on five datasets reveal the effectiveness and efficiency of TLIMSC.

关 键 词：多视图聚类 加权张量核范数 谱嵌入表征 自适应置信度 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]