Title: Multivariate functional subspace classification for high-dimensional longitudinal data and its application
Authors: Tatsuya Fukuda - Chuo University (Japan) [presenting]
Toshihiro Misumi - Yokohama City University (Japan)
Yoshihiko Maesono - Chuo University (Japan)
Sadanori Konishi - Kyushu University (Japan)
Abstract: Classification for high-dimensional longitudinal data with multiple classes plays an important role in diverse fields of the natural and social sciences. The subspace method known as the class-featuring information compression (CLAFIC) based on principal component analysis is a useful tool for classifying and representing patterns, and a number of applications of CLAFIC method have been reported in character recognition, speech recognition, image analysis, etc. However, a disadvantage of this procedure is that it may not be applied to longitudinal studies where subjects are measured at different time points. In order to overcome this issue, we propose a novel classification method for high-dimensional longitudinal data with multiple classes by extending CLAFIC method, and we call it multivariate functional subspace method (mFSM). The mFSM can be used to classify an unlabeled data by measuring the distance between the data and a subspace for each class, obtained by a multivariate functional principal component analysis. Since the accuracy of mFSM based classifier deeply depends on the dimension of subspaces, we consider the problem of selecting the optimal dimension of subspaces. The performance of proposed method is evaluated through a simulation study, and we present the results of the analysis of handwritten number data.