Adjusted Modeling to Deal with Fundamental Comparability Problems in Multiway Data

Models for the analysis of multiway data such as PARAFAC/CANDECOMP implicitly assume that all data entries can be meaningfully compared. However, in data-analytic practice, this assumption is often violated because for at least one data mode the meaning of each of the elements of that mode is not constant across the corresponding data slice. This implies a major problem for the interpretation of the estimated PARAFAC/CANDECOMP model parameters, especially if one hopes these to represent relevant structural characteristics of the elements of the data modes under study. In this paper we will first describe this comparability problem, and clarify it with examples of multiway data in systems biology and psychiatric diagnosis. Next, we will explain how in some cases comparability problems can be successfully addressed by a suitably adjusted type of modeling. In this regard, we will focus on clusterwise PARAFAC/CANDECOMP models, and (ordinary as well as clusterwise) simultaneous component models of suitably unfolded versions of the multiway data. We will illustrate with analyses of sensory profiling data.