I have a term-document frequency matrix which is thus high-dimensional and very sparse. Whenever I generate SOMs in the "Kohonen" package in R I get one node dominating the others no matter which topology or learning rate I use. Is there a clear way of dealing with high-dimensional sparse data because this is effecting the topology of the network in that one node is dominating the rest.
I'm not sure if this answers the question, but I think it might be described in Kohonen's recent book. He suggests to assign a unique random vector to each unique word, each vector as uncorrelated as possible, so that it's the combination of the unique words that "defines the degree of similarity of different local contexts." From p. 119-120 in http://docs.unigrafia.fi/publications/kohonen_teuvo/