Multivariate data analysis: quo vadis?

Abstract

AbstractIndustry and academe are characterized by steadily increasing huge amounts of data with very different data structures. Both static and dynamic data contexts need to be addressed. A new generic, flexible and comprehensive general data‐modelling concept is needed to cope with these demands. During the past 20 years, object‐oriented programming (OOP) has become a de facto industry standard of how programming tasks should be defined and carried out in the context of deterministic data modelling. We present here a first framework of analogous ideas for multivariate data analysis. A new strategy, object‐oriented data modelling (OODM), is proposed which is invariant with respect to the specific data structures and the practical data context. We present a first delineation of meta‐principles, ideas and stimulants for tomorrow's possible development paths of modelling, in which the fundamental data analysis unit is the generalized ‘PLS object’ in the OOP sense. The key novel aspect concerns inter‐object information transfer, facilitated by ‘root‐sum‐of‐squares averaging’ (RSSA), which uses w loading weights as between‐object transfer agents. These features allow a powerful generalization beyond multiblock as well as hierarchical bilinear modelling to be laid out. The present part I outlines a first framework for the new data‐modelling approach, while part II forms a complementing catalogue of specific options and possibilities when implementing the new principles. Copyright © 2003 John Wiley & Sons, Ltd.