Informatics and Applications
2022, Volume 16, Issue 3, pp 52-58
MODEL SELECTION FOR MATRIX FACTORIZATION WITH MISSING COMPONENTS
Abstract
The work is dedicated to the problem of factorizing a matrix with missing components into a product of two lower-rank matrices. The influence of the intensity of missing on the choice of the factorization model is studied. Two algorithms for parameter estimation are considered: alternating least squares (ALS) and Wiberg - for two factorization models: with and without means. It is substantiated that there is no need to use a model with averages: it is a special case of another model and, in some cases, leads to ambiguous solutions. During the experiments, the preference was given to a more stable ALS algorithm. The advantages of the insertion method over random filling in the initial settings of iterative algorithms for estimating model parameters are demonstrated.
The reasons for the negative properties of the existing version of the Wiberg algorithm are revealed. Based on the experiments, it was found that with an increase in the probability of missing, the accuracy of the presentation of the available data increases which leads to an underestimation of the true value of the model dimension.
[+] References (6)
- Chen, P. 2008. Optimization algorithms on subspaces: Revisiting missing data problem in low-rank matrix. Int. J. Comput. Vision 80(1):125-142.
- Hong, J. H., and A. Fitzgibbon. 2015. Secrets of matrix factorization: Approximations, numerics, manifold optimization and random restarts. Available at: http://www.cv- foundation.org/openaccess/content_iccv_2015/papers/ Hong_Secrets_of_Matrix_ICCV_2015_paper.pdf (accessed May 26, 2022).
- Wiberg, T. 1976. Computation of principal components when data are missing. 2nd Symposium of Computational Statistics Proceedings. Wien: Physica-Verlag. 229-236.
- Okatani, T, and K. Deguchi. 2007. On the Wiberg algorithm for matrix factorization in the presence of missing components. Int. J. Comput. Vision 72(3):329-337.
- Krivenko, M. P 2016. Kriterii znachimosti otbora priznakov klassifikatsii [Significance tests of feature selection for classification]. Informatika i ee Primeneniya - Inform. Appl. 10(3):32-40.
- Seber, G.A. F. 2008. A matrix handbook for statisticians. Hoboken, NJ, USA: Wiley. 560 p.
[+] About this article
Title
MODEL SELECTION FOR MATRIX FACTORIZATION WITH MISSING COMPONENTS
Journal
Informatics and Applications
2022, Volume 16, Issue 3, pp 52-58
Cover Date
2022-10-10
DOI
10.14357/19922264220307
Print ISSN
1992-2264
Publisher
Institute of Informatics Problems, Russian Academy of Sciences
Additional Links
Key words
lower-rank matrix approximation; singular value decomposition; missing data; ALS algorithm; Wiberg algorithm
Authors
M. P. Krivenko 
Author Affiliations
Federal Research Center "Computer Science and Control" of the Russian Academy of Sciences, 44-2 Vavilov Str., Moscow 119333, Russian Federation
|