基于超图正则化非负张量链分解的聚类分析  

Clustering Analysis Based on Hyper-graph Regularized Non-Negative Tensor Train Decomposition

作　　者：戴浩磊 黄永慧[1] 周郭许 DAI Haolei;HUANG Yonghui;ZHOU Guoxu(School of Automation,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学自动化学院,广州510006

出　　处：《计算机工程》2023年第6期81-89,共9页Computer Engineering

基　　金：国家自然科学基金(62073087);广东省重点领域研发计划(2019B010154002)。

摘　　要：非负张量链分解作为一种重要的张量分解模型,可保留数据内部结构信息,广泛应用于高维数据的特征提取和表示。从流形学习角度出发,高维数据信息通常潜在于低维空间的非线性流形结构中,然而现有图学习理论只能建模对象间的成对关系,很难准确刻画具有复杂流形结构的高维数据的相似关系。引入超图学习,提出一种超图正则化非负张量链(HGNTT)分解方法,在高维数据中提取低维表示的同时通过构建超图描述样本数据间的高阶关系,从而保留非线性流形结构,同时采用乘法更新方法对HGNTT模型进行优化求解并证明其收敛性。在ORL和Faces95这两个公开数据集上的聚类实验结果表明,相比于NMF、GNMF等方法,HGNTT方法的聚类准确率和归一化互信息分别提升了1.2%~7.6%和0.2%~3.0%,验证了HGNTT方法的有效性。Non-negative Tensor Train(NTT)decomposition,as an important tensor decomposition model,can preserve the internal structure information of data and is widely used in feature extraction and representation tasks of high-dimensional data.From the perspective of manifold learning,high-dimensional data information is usually latent in the nonlinear manifold structure in low-dimensional space.However,existing graph learning theories can only model pairwise relationships between objects,and accurately portraying the similar relationships of high-dimensional data with a complex manifold structure is difficult.By introducing hyper-graph learning,this study proposes a Hyper-Graph regularized Non-negative Tensor Train(HGNTT)decomposition method for extracting low-dimensional representations from high-dimensional data while describing the higher-order relationships between sample data points by constructing hyper-graphs,thereby preserving the nonlinear manifold structure.Moreover,a Multiplicative Update(MU)method is used to optimally solve the HGNTT model and prove its convergence.Clustering experiments on two publicly available datasets,ORL and Faces95,show that the clustering accuracy and Normalized Mutual Information(NMI)of the HGNTT method improves by 1.2%-7.6%and 0.2%-3.0%,respectively,compared with those of NMF and GNMF,thereby validating the effectiveness of the HGNTT method.

关 键 词：非负张量链分解 特征提取 超图学习 乘法更新方法 聚类分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]