Research Article
The Performance of a Kernel-Based Variable Dimension Reduction Method
@INPROCEEDINGS{10.1007/978-3-031-28790-9_4, author={Thanh Do Van and Hai Nguyen Minh}, title={The Performance of a Kernel-Based Variable Dimension Reduction Method}, proceedings={Nature of Computation and Communication. 8th EAI International Conference, ICTCC 2022, Vinh Long, Vietnam, October 27-28, 2022, Proceedings}, proceedings_a={ICTCC}, year={2023}, month={3}, keywords={Time series Big data Nowcast Dimensional reduction Kernel trick Factor model}, doi={10.1007/978-3-031-28790-9_4} }- Thanh Do Van
Hai Nguyen Minh
Year: 2023
The Performance of a Kernel-Based Variable Dimension Reduction Method
ICTCC
Springer
DOI: 10.1007/978-3-031-28790-9_4
Abstract
Building forecast models, especially nowcast models, on large data sets of time series variables is a topic of great interest. The most popular method used to build such models is the dynamic factor model, in which factors are extracted from input data sets using the principal component analysis (PCA) or sparse PCA (SPCA) methods. Many studies have shown that the forecast accuracy of models built under such an approach is higher than that of other benchmark models. But the PCA and SPCA methods are only effective when input data sets approximate a hyperplane, while real-world data sets are not always so.
The purpose of this article is to briefly introduce a kernel-based variable dimension reduction method called the KTPCA, and a process of trial and error of this method called the KTPCA#. Experimenting on real-world data sets at the same sampling frequency as well as mixed sampling frequencies shows that the KTPCA# method is superior to the PCA, SPCA, Randomized SPCA (RSPCA), and robust SPCA (ROBSPCA) methods.
