***** To join INSNA, visit http://www.insna.org *****
Hi all,

The feedback I received about the last email was very helpful. Thank you all for that. I'm back again with another question about the same network.

I've managed to get the method proposed in the paper, "Extracting the multiscale backbone of complex weighted networks", Serrano et al. (2009), PNAS, to reduce enough edges to produce a decent network from my data.

However, some nodes in the network are extremely popular relative to most of the other nodes. This results in the Serrano et al. method leaving upwards of 5,000 edges for the extremely popular nodes, and ~30 or fewer edges for the average nodes. When I cluster this network, it of course results in a giant blob of nodes connected to the extremely popular nodes.

Are there any methods established to handle this node popularity bias? Lowering the alpha significance cutoff doesn't seem to help -- there are always orders of magnitude more edges left for the popular nodes.

I have found that arbitrarily cutting each node down to only having its X strongest edges (where X is some arbitrary small number like 20 or 100) after applying the Serrano et al. method works great and produces a beautiful clustering, but I am doubtful that such a method would stand through peer review.

Any feedback is greatly appreciated. Kind regards,
Randy


On Thu, Nov 7, 2013 at 10:10 PM, Randy Olson <[log in to unmask]> wrote:
Hi all,

I am working with a large, fully-connected social network: 24k nodes, ~24k^2 edges. I have been exploring a number of edge reduction methods to make the network feasible to cluster and visualize, namely the method proposed in the paper, "Extracting the multiscale backbone of complex weighted networks", Serrano et al. (2009), PNAS.

I am finding that the method doesn't reduce enough edges, largely because each node has ~24k edges to begin with. I've tried applying a Bonferroni correction to the alpha cutoff (i.e., alpha = 0.05 / 24k^2), and there were still so many edges remaining that the network was no better than before the edge reduction method. Has anyone had a similar experience with applying the method proposed in the Serrano et al. paper to large networks? What are some possible workarounds?

Cheers,
Randy

--
Randal S. Olson

Computer Science PhD Student
Michigan State University
E-mail: [log in to unmask] | Twitter: @randal_olson



--
Randal S. Olson

Computer Science PhD Student
Michigan State University
E-mail: [log in to unmask] | Twitter: @randal_olson
http://www.randalolson.com
_____________________________________________________________________ SOCNET is a service of INSNA, the professional association for social network researchers (http://www.insna.org). To unsubscribe, send an email message to [log in to unmask] containing the line UNSUBSCRIBE SOCNET in the body of the message.