Linear methods in multimode data analysis for decision making

This paper presents and analyses several methods for the evaluation of information given in the form of many-way matrices. These methods are based on the least squares approximation of a matrix by a many-vector product which can be represented as a nonlinear eigenvector problem. Using real data about university choice by high school graduates in Israel, the authors develop and compare the following three families of methods: parallel proportional profiles, various types of methods based on the use of cyclic matrices (canonical correlations, principal components, and planes’ approximation), and minimization of relative deviations.