IMPROVING THE SPEED OF MULTIWAY ALGORITHMS PART II - COMPRESSION

In this paper an approach is developed for compressing a multiway arra y prior to estimating a multilinear model with the purpose of speeding up the estimation. A method is developed which seems very well-suited for a rich variety of models with optional constraints on the factors . It is based on three key aspects: (1) a fast implementation of a Tuc ker3 algorithm, which serves as the compression method, (2) the optima lity theorem of the CANDELINC model, which ensures that the compressed array preserves the original variation maximally, and (3) a set of gu idelines for how to incorporate optional constraints. The compression approach is tested on two large data sets and shown to speed up the es timation of the model up to 40 times. The developed algorithms can be downloaded from http:\\ newton.mli.kvl.dk \ foodtech.html. (C) 1998 El sevier Science B.V. All rights reserved.