Re: [igraph] 'decompose.graph' versus 'clusters'

[Gabor Csardi]

On Fri, Jul 25, 2008 at 08:13:39AM -0700, David Hunkins wrote:
> Okay, so the 8-way partition did the trick (decompose.graph was able to 
> pull apart the 300k-node graphs but not the 600k-node graphs, which, when 
> run, had generated the protection faults). I think I'm calculating real 
> betweenness values for each connected component, because the order of 
> operations is this:
> 1. remove largest cluster (the 2M-plus node cluster that breaks  
> everything)
> 2. remove smallest clusters (the 1- and 2-node clusters that I'm not  
> interested in)
> 3. take remaining clusters (about 200,000 of them) and divide them up  
> into 8 groups
> 4. for each of the 8 groups, run decompose.graph to return a list of  
> subgraphs
> 5. run betweenness on each of the graphs in the list of subgraphs (so I 
> am only ever running betweenness on something that's a maximal connected 
> component in the original graph)
> If I still haven't understood something about betweenness please let me 
> know.
> This was surprisingly fast (just 8 hours of cpu time).

Oh, ok, that is fine, I thought you wanted to break the giant 
component into eight pieces.....

> Next month I'll be trying such a strategy again on a much larger  
> dataset; I'll be using the faster decompose.graph (presumably that's in 
> your latest 0.6 tarfile) and let you know how it goes.

Not yet! I'll email you when it is uploaded. 

G.

[...]

-- 
Csardi Gabor <address@hidden>    UNIL DGM