A0408
Title: Contrastive principal component analysis in high dimension low sample size
Authors: Shao-Hsuan Wang - National Central University (Taiwan) [presenting]
Kazuyoshi Yata - University of Tsukuba (Japan)
Abstract: Principal Component Analysis (PCA) is a commonly used linear dimensionality reduction method and is often used to visualize a single dataset; Contrastive Component Analysis (cPCA) can be used in situations where there are multiple datasets, and cPCA can explore the unique low-dimensional structure of a specific dataset on the premise of referring to other datasets. However, while cPCA has been shown in many fields to find important data patterns that PCA ignores, cPCA lacks a statistical model to identify why cPCA can identify those changes that are of interest. A statistical model for cPCA is proposed. The target data is divided into the signal matrix that is of interest and the nuisance matrix that is not of interest, and an effort is made to explain that cPCA can remove the influence of the nuisance matrix on the target data. On the other hand, the advantages of cPCA are illustrated in restoring the signal matrix using simulation analysis. Furthermore, a new method is proposed based on the model to help us decide on the contrast parameter that is important to perform cPCA. Finally, data patterns of interest in the synthetic image example are found by adjusting the contrast parameter and verifying that the new method of choosing the contrast parameter can achieve the same effect.
