The N-way Toolbox for MATLAB
Claus A. Andersson (claus@andersson.dk) and
Rasmus Bro (rasmus@optimax.dk)
Royal Veterinary and Agricultural University
Dept. Dairy and Food Science – Food Technology
Rolighedsvej 30
DK-1958 Frederiksberg C
Abstract
This communication describes a free toolbox for MATLAB® for analysis of multi-way data. The toolbox is called “The N-way Toolbox for MATLAB” and is available on the Internet at http://www.models.kvl.dk/source/. This communication is by no means an attempt to summarize the extensive work done in multi-way data analysis but is
intended solely for informing the reader of the existence, functionality, and applicability of the N-way Toolbox for MATLAB.
1. Introduction
The N-way Toolbox for MATLAB® is a freely available collection of functions and algorithms for modelling multi-way data sets by a range of multi-linear models. Several types of models are covered; CANDECOMP-PARAFAC (canonical decomposition - parallel factor analysis), multi-linear PLSR (partial least squares regression),
generalised rank annihilation method (GRAM), direct tri-linear decomposition (DTLD) and the class of Tucker models. When denoting missing observations by NaN (not-a-number), the algorithms apply expectation maximization to obtain the parameters that minimize the least-squares error term.
Selected types of optional constraints have been built into the least-squares error minimisation algorithms for CANDECOMP-PARAFAC and Tucker models; non-
negativity, uni-modality, and orthogonality. Different constraints may be set up for the different modes. In addition to these constraints, the structure of the Tucker models can be forced to allow only selected factor interactions. Furthermore, three methods for core simplification by orthogonal rotations have been implemented. Most algorithms can handle any number of modes (N≥ 3) in data.
The N-way Toolbox for MATLAB can be downloaded via Internet from
http://www.models.kvl.dk/source/. Two interactive Internet courses accompany The N-way Toolbox for MATLAB, and they are available at
http://www.models.kvl.dk/courses/. Both the Tucker and the PARAFAC courses come with real and simulated multi-way data sets and are intended for training in applying the models to different kinds of problems.