| 本文已被：浏览 3948次 下载 6417次 |
码上扫一扫！ |
|
|
| 面向时序数据的矩阵分解 |
|
黄晓宇1,2, 潘嵘2, 李磊2, 梁冰3, 陈康3, 蔡文学1
|
|
1.华南理工大学 经济与贸易学院, 广东 广州 510006;2.中山大学 计算机软件研究所, 广东 广州 510275;3.中国电信股份有限公司 广东研究院, 广东 广州 510630
|
|
| 摘要: |
| 研究一类特殊的矩阵分解问题:对由多个对象在一组连续时间点上产生的数据构成的矩阵R,寻求把它近似地分解为两个低秩矩阵U和V的乘积,即R≈UT×V.有为数众多的时间序列分析问题都可归结为所研究问题的求解,如金融数据矩阵的因子分析、缺失交通流数据的估计等.提出了该问题的概率图模型,进而由此导出了其约束优化模型,最终给出了模型的求解算法.在不同的数据集上进行实验验证了该模型的有效性. |
| 关键词: 矩阵分解 时间序列数据 概率图模型 缺失估计 低秩近似 |
| DOI：10.13328/j.cnki.jos.004718 |
| 分类号: |
| 基金项目:国家高技术研究发展计划(863)(2012AA12A203); 国家自然科学基金(61003140); 国家社会科学基金(13BTJ005) |
|
| Matrix Factorization for Time Series Data |
|
HUANG Xiao-Yu1,2, PAN Rong2, LI Lei2, LIANG Bing3, CHEN Kang3, CAI Wen-Xue1
|
|
1.School of Economics and Commerce, South China University of Technology, Guangzhou 510006, China;2.Software Institute, Sun Yet-San University, Guangzhou 510275, China;3.Academy of Guangdong Telecom Company, China Telecom Corporation Limited, Guangzhou 510630, China
|
| Abstract: |
| The paper studies a matrix factorization problem for time series data, where the target matrix R consists of the equal length time series data generated by a set of objects. The goal is to find two low rank matrices U and V, such that R≈UT×V. Many time series analysis problems, such as finance data analysis and missing traffic data imputation, can be reduced to the proposed model. A probabilistic graphical representation for the problem is proposed, and a constrained optimization model from the graphical representation is derived. The solution algorithms for the proposed model is also presented. Empirical studies show that the proposed model is superior to the baselines. |
| Key words: matrix factorization time series data probabilistic graphical model missing estimation low rank approximation |
