基于低秩张量图学习的不完整多视角聚类  

Low-rank Tensor Graph Learning Based Incomplete Multi-view Clustering

作　　者：文杰 颜珂 张正 徐勇 WEN Jie;YAN Ke;ZHANG Zheng;XU Yong(Shenzhen Key Laboratory of Visual Object Detection and Recognition,Harbin Institute of Technology(Shenzhen),Shenzhen 518055;School of Computer Science and Technology,Beijing Institute of Technology,Beijing 100080;Peng Cheng Laboratory,Shenzhen 518055)

机构地区：[1]哈尔滨工业大学(深圳)深圳市视觉目标检测与判识重点实验室,深圳518055 [2]北京理工大学计算机科学与技术学院,北京100080 [3]鹏城实验室,深圳518055

出　　处：《自动化学报》2023年第7期1433-1445,共13页Acta Automatica Sinica

基　　金：深圳市基础研究项目(JCYJ20190806142416685);国家自然科学基金(62006059,62002085);中国博士后科学基金(2020M681099);中国博士后创新人才支持计划(BX20190100);深圳市科技创新委员会重点实验室组建项目(ZDSYS20190902093015527)资助。

摘　　要：传统多视角聚类都基于视角完备假设,要求所有样本的视角信息完整,不能处理存在部分视角缺失情形下的不完整多视角聚类任务.为解决该问题,提出一种基于低秩张量图学习的不完整多视角聚类方法.为了恢复相似图中缺失视角所对应的样本关联信息,该方法将低秩张量图约束和视角内在图保持约束融入到多视角谱聚类模型.通过在一个统一模型中同时挖掘视角间的互补信息和视角内未缺失样例的关联信息,所提出的方法能够得到表征样例邻接关系的完整相似图和视角间一致的最优聚类指示矩阵.与12种不完整多视角聚类方法进行实验对比,实验结果表明所提出的方法在多种视角缺失率下的5个数据集上获得了最好的聚类性能.Conventional multi-view clustering methods all assume that the give multi-view data is complete,i.e.,all views are fully observed,which are not applicable to the incomplete multi-view clustering case with missing-views.To address this issue,we propose a method,called low-rank tensor graph learning(LASAR).To recover the missing connections corresponding to the missing views in the graph space,the proposed method integrates the low-rank tensor graph constraint and intra-view graph constraint into the multi-view spectral clustering framework.By exploring the inter-view information and intra-view information with respect to the observed views simultaneously,the proposed method can obtain the optimal completed graphs of all views and the optimal clustering indicator matrix shared by all views.Experimental results on five datasets with different missing-view rates show that the proposed method obtains better performance than 12 state-of-the-art incomplete multi-view clustering methods.

关 键 词：多视角聚类 视角缺失 不完整多视角聚类 图学习 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]