Self-organising map was learned using the SOM Toolbox1 for Matlab. I implemented the soft reconstruction with Matlab. The number of model vectors and the width of the softening kernels associated with them were left as parameters and the selection of values is not studied herein. Instead, the values with the best results were used. Number of iterations was the default value in the SOM toolbox and additional 50 iterations through all the data were run with neighbourhood set off.
I modified the NFA code for Matlab to support missing values. The code needed to reproduce the experiments will be publicly available2. The NFA method was tried with fixed number of hidden neurons using several different number of factors only, because the algorithm requires about ten thousand batch iterations, where one batch iteration means going through all the observations once.