Title: Multivariate functional subspace methods for classifying high-dimensional longitudinal data
Authors: Tatsuya Fukuda - Chuo University (Japan) [presenting]
Toshihiro Misumi - Yokohama City University (Japan)
Hidetoshi Matsui - Shiga University (Japan)
Sadanori Konishi - Chuo University (Japan)
Abstract: Classification of high-dimensional longitudinal data with multiple classes plays an important role in various fields of science, such as medical research, meteorology and ecology, and those data are often measured at different time points for individual. Existing approaches for classifying multivariate observations are mainly restricted to data measured at same time points. We propose a novel multi-class classification procedure for high-dimensional longitudinal data based on CLAss-Featuring Information Compression (CLAFIC) method with the help of multivariate functional principal component analysis. We call this method multivariate functional subspace method. The multivariate functional subspace method can be used to classify unlabeled data according to the distance between the data and a subspace for each class, obtained by a multivariate functional principal component analysis. In modeling process of longitudinal observations to functional data, we use a smoothing technique via regularized basis expansions with Bernstein polynomials. We examine the performance of this method through the analysis of a real data set and a simulation study.