计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (7): 1251-1260.DOI: 10.3778/j.issn.1673-9418.1905065

• 图形图像 • 上一篇    

鲁棒概率矩阵三分解

史加荣,陈姣姣   

  1. 1. 省部共建西部绿色建筑国家重点实验室/西安建筑科技大学,西安 710055
    2. 西安建筑科技大学 理学院,西安 710055
  • 出版日期:2020-07-01 发布日期:2020-08-12

Robust Probabilistic Matrix Tri-factorization

SHI Jiarong, CHEN Jiaojiao   

  1. 1. State Key Laboratory of Green Building in Western China, Xi'an University of Architecture and Technology, Xi'an 710055, China
    2. School of Science, Xi'an University of Architecture and Technology, Xi'an 710055, China
  • Online:2020-07-01 Published:2020-08-12

摘要:

矩阵分解是计算机视觉、机器学习和数据挖掘中经常使用的数据分析工具。近年来,矩阵分解的概率模型已成为人们关注的焦点。现有的概率矩阵分解一般将数据矩阵分解为两个低秩矩阵之积,这可能会限制该模型的灵活性和实用性。为此,提出了鲁棒概率矩阵三分解模型(RPMTF)。该模型将数据矩阵分解为三个矩阵的乘积,并考虑了其鲁棒性。在模型求解时,先将拉普拉斯分布进行分层表示;再采用基于极大后验估计的策略,设计了一种条件期望最大化算法。在实验中,将鲁棒概率矩阵三分解应用到图像去噪和视频背景建模中,结果证实了所提方法的可行性与有效性。

关键词: 矩阵三分解, 概率矩阵分解, 期望最大化, 极大后验估计

Abstract:

Matrix factorization is a commonly-used data analysis tool in computer vision, machine learning and data mining. In recent years, the probabilistic models of matrix factorization have become the focus of attention. Existing probabilistic matrix factorization models generally decompose a given data matrix into the product of two low rank matrices, which probably limits the flexibility and practicability of these models. To address this issue, this paper proposes a model of robust probabilistic matrix tri-factorization (RPMTF). This model factorizes the data matrix    into the product of three matrices and the robustness is taken into account simultaneously. A hierarchical representation of the Laplace distribution is firstly adopted for solving the proposed model. Then an expectation maximization algorithm is designed based on a strategy of maximum a posterior estimation. In the experiment, RPMTF is applied to image denoising and video background modeling, and the results verify the feasibility and effectiveness of the proposed method.

Key words: matrix tri-factorization, probabilistic matrix factorization, expectation maximization, maximum a posterior estimation