Re: [igraph] bug in Graph.community_leading_eigenvector?

[Gabor Csardi]

Furthermore, 1) you cannot set the number of iterations in the old
implementation (except for changing the source), so it'll be hard to
compare, 2) two different methods are used for eigenvector
calculations, there can be differences in the results id the first two
eigenvalues are close to each other, or 3) the 0.4.5 version might be
simply wrong. 

Anyway, if you want to compare the results, then you can calculate 
them "by hand" for a small matrix. Or you can take a look at the 
'inetwork' R package, that has an implementation too. It is quite 
slow, but should be ok for small graphs. I'm pretty sure the 
0.5 version is good. 

Actually, if you want to optimize modularity, the fastgreedy and
especially the spinclass methods tend to work better. 

Gabor

On Mon, Mar 24, 2008 at 05:31:54PM +0100, Tamas Nepusz wrote:
> Hi,
> 
> >Using igraph 0.5 in python 2.5 in osx 10.4, i'm running the  
> >leading_eigenvector community detection method and I get the  
> >following warning:
> Eigenvector calculations in igraph 0.4.5 were not based on ARPACK, so  
> that's the cause why there's no warning in 0.4.5. I can't judge  
> whether the result of 0.4.5 is correct or the result of 0.5; but maybe  
> raising the maximum number of iterations allowed in the ARPACK module  
> will help. The Python interface has a variable called arpack_options,  
> which is an instance of class ARPACKOptions (see  
> help(arpack_options)). arpack_option has an attribute called maxiter  
> (which is 3000 per default) that can be modified to allow more  
> iterations. Maybe this helps, maybe not.
> 
> -- 
> Tamas
> 
> 
> 
> _______________________________________________
> igraph-help mailing list
> address@hidden
> http://lists.nongnu.org/mailman/listinfo/igraph-help

-- 
Csardi Gabor <address@hidden>    UNIL DGM