基于张量学习的潜在多视图子空间聚类  

Tensor Learning-based for Latent Multi-view Subspace Clustering

作　　者：李理[1,2] 李敬豪[1,2] 张小乾 LI Li;LI Jinghao;ZHANG Xiaoqian(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China;Advanced Control and Modeling Laboratory,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621010 [2]西南科技大学先进控制与建模实验室,四川绵阳621010

出　　处：《西南科技大学学报》2022年第3期52-59,共8页Journal of Southwest University of Science and Technology

基　　金：国家自然科学基金(62102331)。

摘　　要：现有的基于张量的多视图子空间聚类方法在学习过程中仅考虑张量的全局低秩结构,忽略了表示矩阵的局部信息,且这些方法因数据维度高易受到噪声干扰。提出了一种基于张量学习的潜在多视图子空间聚类方法(TLLMSC)。将高维多视图数据投影到低维嵌入空间中,以便学习到干净字典,消除冗余信息及噪声对聚类性能的影响;将低秩投影融入到基于张量学习的多视图子空间聚类框架中,充分挖掘多视图数据的高阶信息;引入拉普拉斯秩约束直接对数据进行分类,提升TLLMSC的聚类效率;设计了一种基于交替方向乘子法的优化算法对TLLMSC进行高效求解。在BBCsport数据集和Washington数据集上进行的聚类实验结果表明,所提方法的聚类性能优于现有先进的多视图聚类方法。The existing tensor-based multi-view subspace clustering methods only consider the global low-rank structure of tensors in the learning process,ignoring the local information of the representation matrix,and these methods are susceptible to noise interference due to high data dimensions.A tensor learning-based for latent multi-view subspace clustering(TLLMSC)method is proposed.TLLMSC firstly projects high-dimensional multi-view data into a low-dimensional embedding space in order to learn a clean dictionary and eliminate the influence of redundant information and noise on clustering performance.Secondly,incorporating low-rank projection into the multi-view subspace clustering framework based on tensor learning can fully mine the high-order information of multi-view data.Finally,the Laplacian rank constraint is introduced,which classifies the data and improve the clustering efficiency of TLLMSC.In addition,an optimization algorithm based on the alternating direction method of multipliers(ADMM)is designed to efficiently solve the TLLMSC.The clustering experiment results on BBCsport dataset and Washington dataset show that the clustering performance of the proposed method outperforms the existing advanced multi-view clustering methods.

关 键 词：多视图子空间聚类 张量学习 低秩投影 拉普拉斯秩约束 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]