紀要論文

深層展開を用いた行列補完手法の高速化

An Acceleration Technique for Matrix Completion using Deep Unfolding

jpn

24321192

https://doi.org/10.15002/00026289

法政大学大学院情報科学研究科

法政大学大学院紀要. 情報科学研究科編

18

1

4

2023-03-24

Version of Record

This paper proposes an acceleration technique using deep unfolding for matrix completion problem, which is the problem of estimating the missing values of a matrix. To solve the problem, various methods have been proposed, which is based on matrix rank minimization. While these methods have iterative schemes to minimize the matrix rank on each step, they have poor estimation accuracy and convergence property when the parameters of the iterative schemes are not given appropriately. In order to provide appropriate parameters of iterative schemes, this paper focuses on a deep unfolding to accelerate matrix completion methods. The deep unfolding is a method of constructing more flexible derivation algorithms by embedding learnable parameters into existing iterative algorithms and by learning the parameters applying deep learning techniques to differentiable iterative algorithms with inputs and outputs. Algorithms which have adjustable parameters such as gradient descent are expected to accelerate by using deep unfolding. This paper proposes two matrix completion methods accelerated by deep unfolding which can optimize the iterative parameters via machine learning. Numerical experiments show the effectiveness of the proposed methods.

Article

AA12746425