基于结构化张量学习的多视图聚类  

Multi-view clustering based on structured tensor learning

作　　者：李心雨 康可涵 彭冲 Li Xinyu;Kang Kehan;Peng Chong(School of Computer Science&Technology,Qingdao University,Qingdao Shandong 266071,China;School of Computer Science&Techno-logy,Ocean University of China,Qingdao Shandong 266100,China)

机构地区：[1]青岛大学计算机科学技术学院,山东青岛266071 [2]中国海洋大学计算机科学与技术学院,山东青岛266100

出　　处：《计算机应用研究》2025年第2期448-454,共7页Application Research of Computers

基　　金：山东省高等学校青年创新团队资助项目(2022KJ149)。

摘　　要：多视图聚类方法随着数据获取途径日益多样化成为研究热点,但大多数聚类方法低估了噪声和数据多结构互补性信息对聚类结果的影响,并且忽略了聚类结果对低秩张量优化过程的反向引导作用。为解决这些问题,提出了基于结构化张量学习的多视图聚类(multi-view clustering based on structured tensor learning,MCSTL)。首先,对初始表示张量进行再次去噪使其更具准确性和鲁棒性;同时,互补地学习局部结构、全局结构和各视图间的高阶相关性,提高表示张量与原始数据本质簇结构的一致性;然后,从跨视图信息融合的亲和矩阵中学习到统一的特征矩阵,利用其隐含的聚类结构信息反向引导表示张量的优化过程;最后,对特征矩阵施加了正交约束,使其提供数据的软标签信息,并对模型进行直接聚类解释。实验表明,MCSTL在6种聚类评价指标上均表现优异,30个指标数据中有27个达到最优,从而充分验证了MCSTL的有效性和优越性。Multi view clustering methods have become a research hotspot with the increasing diversity of data acquisition techniques.However,most clustering methods underestimate the impact of noise and complementary structural information of the data.Moreover,they often ignore the reverse guidance of clustering results on the optimization process of low rank tensors.To address these issues,this paper proposed a multi-view clustering method based on structured tensor learning(MCSTL).First,it further denoised the initial representation tensor to enhance its accuracy and robustness.At the same time,it complementarily learnt local structure,global structure,and high-order correlation across different views,which improved the consistency between the representation tensor and the intrinsic cluster structure of the original data.Then,it learnt a unified feature matrix from the affinity matrix of cross-view information fusion,and utilized the implicit clustering structure information within it to inversely guide the optimization process of the representation tensor.Lastly,it imposed an orthogonal constraint on the feature matrix,which provided soft label information of the data and allows for a direct clustering interpretation of the model.The experimental results show the MCSTL performs well in all six clustering evaluation metrics,with 27 out of 30 metrics reaching the optimal level,fully verifying the effectiveness and superiority of the MCSTL method.

关 键 词：多视图聚类 张量 结构性约束 特征矩阵 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]