Details of Research Outputs

TitleA REGULARIZATION/GAUSSIAN PROCESS APPROACH TO SPATIAL-TEMPORAL DATA MODELING
Author (Name in English or Pinyin)
KUANG, Ye
Author (Name in Chinese)邝也
Degree TypeMaster of Philosophy
Supervisor(s) (Name in English or Pinyin)CHEN, Tianshi
Date Issued2019
Degree GrantorThe Chinese University of Hong Kong
Place of ConferralShenzhen
Degree DisciplineComputer and Information Engineering
languageEnglish
Abstract

Spatial-temporal data are widely used in many areas, such as weather, wind and ocean currents. We consider the spatial-temporal data modeling problem as a function estimation problem. The function of the ‘true’ model can be viewed as a Gaussian process. Gaussian process regression is a standard approach to estimating function values for the given inputs. We try two implementations of Gaussian process regression, a straightforward one and an alternative one via Kalman filter and smoother. Gaussian process regression via Kalman filter and smoother with computational complexity O(NM3) is more efficient than the straightforward one with computational complexity O(N3M3), where N is the number of time instants in the training set, M is the number of spatial locations in the training set. We also try several methods to estimate hyper-parameter in the kernel functions of the Gaussian process, including marginal likelihood, generalized cross validation and Stein’s unbiased risk estimation. The alternative implementations of estimation methods use properties of innovations sequence in Kalman filter. The alternative implementations with computational complexity O(NM3) is more efficient than the straightforward implementations with computational complexity O(N3M3), while the accuracy is the same. 时空数据是一种常见于多个领域的数据,如天气,风和洋流。我们将时空数据建模问题视为函数估计问题。‘真实’ 模型的函数可以看作是高斯过程。高斯过程回归是估计给定输入处的函数值的常用方法。我们将尝试两种高斯过程回归的实现方法。一种是高斯过程回归的直接实现,另一种是通过卡尔曼滤波和平滑的替代实现。在准确度一样的前提下,通过卡尔曼滤波和平滑的高斯过程回归的计算复杂度为O(NM3),直接的高斯过程回归的计算复杂度O(N3M3)。通过卡尔曼滤波和平滑的高斯过程回归的计算复杂度更低。其中N 是训练集中观测时刻的个数,M 是训练集中地点的个数。同时,我们还尝试了三种种方法用于估计高斯过程的核函数中的超参数。三种方法分别是:边缘似然、广义交叉验证与Stein 无偏风险估计。对于这三种参数估计方法,我们都可以通过卡尔曼滤波中残差序列的性质得到便捷的实现方法。在准确度一样的情况下,通过卡尔曼滤波中残差序列的性质得到的实现方法的计算复杂度为O(NM3)。此方法相较于计算复杂度为O(N3M3) 的直接实现方法更高效。

LibraryUniversity Library
Location Theses & Dissertations Collection
Call NumberM.Phil. K83 2019
Document TypeThesis
Identifierhttps://irepository.cuhk.edu.cn/handle/3EPUXD0A/2719
LinksPRIMO
CollectionSchool of Science and Engineering
Recommended Citation
GB/T 7714
KUANG, Ye. A REGULARIZATION/GAUSSIAN PROCESS APPROACH TO SPATIAL-TEMPORAL DATA MODELING[D]. The Chinese University of Hong Kong, Shenzhen,2019.
Files in This Item:
File Name/Size DocType File Type Version Access License
Thesis-邝也.pdf(990KB)Thesis-- No AccessCC BY-NC-SA
Related Services
Usage statistics
Google Scholar
Similar articles in Google Scholar
[KUANG, Ye]'s Articles
Baidu academic
Similar articles in Baidu academic
[KUANG, Ye]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[KUANG, Ye]'s Articles
Terms of Use
No data!
Social Bookmark/Share
Please consult the service desk at the University Library regarding access of the hard copy.
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.