I'm working with some graph matching algorithms that output a similarity matrix between nodes from different graphs and that I wish to match as best as possible. Unfortunately, some nodes might be too uninformative and match well with a lot of nodes and that will affect the matching of nodes that are informative and might have just one strong match. I've read that applying the Sinkhorn Knopp would normalize and make row sums equal to 1 and columns sums equal to 1 too and in the process decrease the influence of uninformative nodes and increase the effect of informative ones.
As I'm quite unfamiliar with Mathematics and can't find any good explanation, I'm asking why would this transformation have the effect I want?
Also, I'll rarely have a square matrix so I think that bistochastic normalization is out of the way? I've use l2-norm and the rows and cols sums are relatively balanced and all close to 1. Will l2-norm suffice for what I want?
Thanks in advance.