Simultaneous simplification of loading and core matrices in N-way PCA: application to chemometric data arrays

Abstract

In the Tucker3 model of N-way principal components analysis (NPCA), a so-called core matrix describes the possible interactions between components from different modes. For an easy interpretation of solutions, it is necessary to have as few interactions as possible (in conventional PCA of data tables, such interactions can always be avoided). This goal may be realized by various approaches of core matrix transformations. At the same time, it is desirable to have simple component (or loading) matrices. Usually, the simplicity of the core conflicts to a certain degree with the simplicity of the components. The paper demonstrates how the conditional optimization of both goals can be used to find a compromise. For the purpose of illustration, the procedure is first applied to a small three-way data array from heavy metal analysis of tissues in different samples of game. Later, a data array of bigger size from a three-way interlaboratory study is considered.