Re: [igraph] Modularity Methods in iGraph (python)

[Tamas Nepusz]

Hi,

I am finding ways of divide my network into modules and I saw that the package in python comes with several alternatives. I wanted to implement the Modularity (Q) as defined by Newman on 2006

"Modularity" is only a measure that tries to quantify how "good" a partition is; Newman's paper defines one possible method to maximize modularity (heuristically), but it is only one of many possible methods, some of which work better in certain types of networks and perform worse on others. That's why igraph implements multiple possible heuristics for finding modules in a network.

 

It seems that the same formula shown on PNAS was used in the “community_leading_eigenvector” method, that is based on https://arxiv.org/pdf/physics/0605087.pdf. Is this correct?

Yes, I believe this is the same method.

 

Also, for large sparse matrices, would you rather recommend this method or the “community_fastgreedy” optimization?

Well, I don't recommend any particular method because all these methods have their strengths and weaknesses, and there are entire papers dedicated to the meta-analysis of all the methods that have been proposed in the literature for modularity maximization (see, e.g., https://arxiv.org/abs/0906.0612, which is 103 pages on its own). Personally, I usually try "community_multilevel" first, but that's just a matter of taste. I also like "community_infomap" because it has a scientifically solid theoretical background. The advantage of "community_infomap" is that it does not try to maximize modularity itself but a different measure; the reason why it is an advantage is because you don't fall into the trap of circular reasoning (which you would do by using a method to find a clustering that has a high modularity, and then claiming that your findings are relevant _because_ the clustering has a high modularity).

T.