Fast Projection Singular Value Thresholding for Low Rank Optimization and Application
Download as PDF
DOI: 10.23977/cii2019.15
Author(s)
Shan Wang, Jianhui Hu, Chuyi Xu
Corresponding Author
Shan Wang
ABSTRACT
Recovering an unknown matrix from corrupted observations is known as the matrix completion problem, it is fundamental to a number of tasks. However, existing most algorithms such as Singular Value Thresholding (SVT) heavily depend on the initialization, which will bring large computational complexity. In this paper, we propose a fast projection singular value thresholding (FPSVT) method, with which we can accelerate the iteration. The key idea is using a projection operator to get an improved initialization which is closer to the unknown optimal solution and using an adaptive thresholding in each iteration of our algorithm. We demonstrate the utility of the proposed method in numerical simulations. The experimental result gives empirical evidence on efficient improvements of the proposed algorithm.
KEYWORDS
Matrix Completion, fast projection singular value thresholding (FPSVT), adaptive thresholding, projection operator