Convergence of a Fast Hierarchical Alternating Least Squares Algorithm for Nonnegative Matrix Factorization

被引:1
|
作者
Hou, Liangshao [1 ]
Chu, Delin [2 ]
Liao, Li-Zhi [1 ]
机构
[1] Hong Kong Baptist Univ, Dept Math, Kowloon Tong, Hong Kong, Peoples R China
[2] Natl Univ Singapore, Dept Math, Singapore 119076, Singapore
关键词
Convergence; Power line communications; Signal processing algorithms; Linear programming; Approximation algorithms; Roads; Postal services; Kurdyka-Lojasiewicz property; nonnegative matrix factorization; the fast hierarchical alternating least squares algorithm; OPTIMIZATION; MINIMIZATION;
D O I
10.1109/TKDE.2023.3279369
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
The hierarchical alternating least squares (HALS) algorithms are powerful tools for nonnegative matrix factorization (NMF), among which the Fast-HALS, proposed in [A. Cichocki and A.-H. Phan, 2009], is one of the most efficient. This paper investigates the convergence of Fast-HALS. First, a more general weak convergence (converged subsequences exist and converge to the stationary point set) is established without any assumption, while most existing results assume all the columns of iterates are strictly away from the origin. Then, a simplified strong convergence (the entire sequence converges to a stationary point) proof is provided. The existing strong convergence is attributed to the block prox-linear (BPL) method, which is a more general framework including Fast-HALS as a special case. So, the convergence proof under BPL is quite complex. Our simplified proof explores the structure of Fast-HALS and can be regarded as a complement to the results under BPL. In addition, some numerical verifications are presented.
引用
收藏
页码:77 / 89
页数:13
相关论文
共 50 条
  • [21] HIERARCHICAL ALTERNATING LEAST SQUARES ALGORITHM WITH SPARSITY CONSTRAINT FOR HYPERSPECTRAL UNMIXING
    Jia, Sen
    Qian, Yuntao
    Li, Jiming
    Li, Yan
    Ming, Zhong
    2010 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING, 2010, : 2305 - 2308
  • [22] PARTITIONED HIERARCHICAL ALTERNATING LEAST SQUARES ALGORITHM FOR CP TENSOR DECOMPOSITION
    Anh-Huy Phan
    Tichavsky, Petr
    Cichocki, Andrzej
    2017 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), 2017, : 2542 - 2546
  • [23] Fast Rank-2 Nonnegative Matrix Factorization for Hierarchical Document Clustering
    Kuang, Da
    Park, Haesun
    19TH ACM SIGKDD INTERNATIONAL CONFERENCE ON KNOWLEDGE DISCOVERY AND DATA MINING (KDD'13), 2013, : 739 - 747
  • [24] A fast algorithm for hyperspectral unmixing based on constrained nonnegative matrix factorization
    Liu, Jian-Jun
    Wu, Ze-Bin
    Wei, Zhi-Hui
    Xiao, Liang
    Sun, Le
    Tien Tzu Hsueh Pao/Acta Electronica Sinica, 2013, 41 (03): : 432 - 437
  • [25] Large margin based nonnegative matrix factorization and partial least squares regression for face recognition
    Pan, Ji-Yuan
    Zhang, Jiang-She
    PATTERN RECOGNITION LETTERS, 2011, 32 (14) : 1822 - 1835
  • [26] Solving non-negative matrix factorization by alternating least squares with a modified strategy
    Hongwei Liu
    Xiangli Li
    Xiuyun Zheng
    Data Mining and Knowledge Discovery, 2013, 26 : 435 - 451
  • [27] Solving non-negative matrix factorization by alternating least squares with a modified strategy
    Liu, Hongwei
    Li, Xiangli
    Zheng, Xiuyun
    DATA MINING AND KNOWLEDGE DISCOVERY, 2013, 26 (03) : 435 - 451
  • [28] Alternating Iteratively Reweighted Least Squares Minimization for Low-Rank Matrix Factorization
    Giampouras, Paris V.
    Rontogiannis, Athanasios A.
    Koutroumbas, Konstantinos D.
    IEEE TRANSACTIONS ON SIGNAL PROCESSING, 2019, 67 (02) : 490 - 503
  • [29] Regularized alternating least squares algorithms for non-negative matrix/tensor factorization
    Cichocki, Andrzej
    Zdunek, Rafal
    ADVANCES IN NEURAL NETWORKS - ISNN 2007, PT 3, PROCEEDINGS, 2007, 4493 : 793 - +
  • [30] Low-Complexity Recursive-Least-Squares-Based Online Nonnegative Matrix Factorization Algorithm for Audio Source Separation
    Lee, Seokjin
    IEICE TRANSACTIONS ON INFORMATION AND SYSTEMS, 2017, E100D (05): : 1152 - 1156