一种矩阵分解和相似度矩阵学习的多视图聚类算法  

A Multiview Clustering Algorithm Based on Matrix Factorization and Similarity Matrix Learning

作　　者：王胜 苏艳苹[1] WANG Sheng;SU Yanping(School of Intelligent Engineer,Zhengzhou University of Aeronautics,Zhengzhou 450046,China)

机构地区：[1]郑州航空工业管理学院智能工程学院,河南郑州457500

出　　处：《郑州航空工业管理学院学报》2023年第5期64-69,86,共7页Journal of Zhengzhou University of Aeronautics

基　　金：河南省科技攻关项目(202102210334)。

摘　　要：考虑到多视图数据之间的互补性和每个视图数据的高维度特性,提出了一种矩阵分解和相似度矩阵学习的多视图聚类算法。为了去除数据中存在的噪声,通过对每个视图的数据进行矩阵分解得到其潜在表示。最大化不同视图表示之间的独立性来得到不同视图的互补信息。为了融合不同视图的潜在表示,最大化最终数据表示的相似度与潜在表示的相似度。采用了最大化熵正则来限制潜在相似度矩阵的值和各视图的权重。3个真实数据库的实验表明,与比较算法相比,本文的聚类算法准确率分别高15%、9%和25%。Considering the complementarity between multiview data and the high-dimensional characteristics of each view,a multiview clustering algorithm based on matrix factorization and similarity matrix learning is proposed.To eliminate the noise in the data,the latent representation of each view is obtained by matrix decomposition..The complementary information of different views is obtained by maximizing the independence between different view representations.In order to integrate representations of multiview data,the similarity between the final data representation and the latent representation is maximized.The maximum entropy regularization is adopted to constrain the values of the latent similarity matrix and the weight of each view.Experiments on three real databases demonstrate that the proposed algorithm achieves clustering accuracy improvements of 15%,9%,and 25%respectively,compared to the comparative algorithms.

关 键 词：矩阵分解 多视图学习 相似度学习 聚类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]