Re: [igraph] use of community detection functions

[Tamas Nepusz]

Hi Simone,

> I should get 2 partitions and I suppose a higher modularity value.
No, you shouldn't necessarily get two partitions. I mean, the fact  
that the Zachary karate club network was split into two parts during  
the study of Zachary does not necessarily mean that it _actually_ has  
two partitions and it attains the highest modularity with two  
partitions. E.g., there is a small separated subgroup of 5 people that  
are connected to the rest of the network only via node 1, this  
subgroup frequently becomes another community according to many other  
methods.

I just checked the Zachary karate club network with the original  
implementation of the fast greedy algorithm of Clauset. It found  
exactly the same partition as igraph with the same number of  
communities: the highest modularity was attained with three  
communities and Q=0.3806, while the modularity was slightly lower for  
two communities: Q=0.3718.

The implementation of the walktrap community detection algorithm is  
exactly the same what the authors used in their paper (actually they  
contributed their code to igraph, Gabor just made a wrapper around  
their code), so this one should also give the same results as the  
original implementation.

Note that the walktrap community detection method can attain higher  
modularity when invoked with random walks of length 3 instead of 4  
(the default):

In [1]: g=load("zachary.graphml")
In [2]: cl=g.community_walktrap()
In [3]: cl.q
Out[3]: 0.35322156548500061
In [4]: cl=g.community_walktrap(steps=3)
In [5]: cl.q
Out[5]: 0.41978961229324341
In [6]: len(cl)
Out[6]: 4

--
T.