Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Consider large (100,000+ vertices, say) graphs, which we think of as representing some population with edges representing some form of symmetric relation. They might be the Friend graph of Facebook, mathematicians with the collaboration relation, or a large computer network.

These networks have the property that they are neither highly structured, nor totally random. No information about other edges on the graph can tell me for certain whether a given pair of vertices is connected. That said, if a given pair of vertices have many common neighbors, then it is considerably more likely that they are connected by an edge (so it is not entirely random). I've seen some lectures on graphs like this, and I understand they are a productive area of research (see, for instance, Kleinberg or Lovasz).

I am curious about the following phenomenon (my description is vague, but part of my question is asking for a good definition). These networks tend to have subsets (which I will call 'clumps') which are significantly more connected to each other than to the average vertex in the graph. Consider a college in Facebook or a research group in mathematics, for example. If the graphs were small enough to draw in a reasonable way, such clumps would be obvious to the naked eye. For very large graphs, this is impractical; so instead, I ask,

1) What is a graph-theoretic way to characterize these clumps?

Clearly, there won't be a yes-no criterion, but I am hoping for some quantity that measures how much a given subset is a clump. This should also factor in the statistical significance of the clump. Very small subsets which are highly connected will happen even in totally random graphs, whereas a large subset which is even moderately well-connected is unlikely in a random graph, and would be interesting to find.

2) Given a graph (and a definition of a clump), how does one find the clumps?

Is there a definition and an algorithm so robust that it can take networks like Facebook or the collaboration graph, and return the clumps that we know are there, like colleges or research discplines?


Oh, and I am not looking for the Szemeredi partition of a graph, which has some similarities to the kinds of partitions I am looking for, but is explicitly a partition of the graph into similar sized chunks. The clumps in a graph don't have to be the same size, disjoint, or contain every element.

share|improve this question
    
It sounds to me like you are looking for induced subgraphs with high edge connectivity. See: en.wikipedia.org/wiki/Connectivity_%28graph_theory%29 –  Harry Stern Dec 9 '10 at 6:50
3  
Related: Clustering coefficient? Also, graph clique might be relevant. –  user4143 Dec 9 '10 at 7:37
    
Clustering coefficients are definitely attacking the kind of problem I am interested in, but they seem to be oddly focused on the correlation between 1-step paths and 2-step paths. They would, for instance, miss a chunk of the graph which looked like a square grid. More generally, if there was a large clump which was only slightly more connected than the general graph, then the difference in clustering coefficients could be non-existent, yet the clump would still be statisically significant (that is, highly unlikely to have randomly occured). –  Greg Muller Dec 9 '10 at 15:57
    
Oh and the Barabási-Albert and Watts and Strogatz models are like your problem in reverse, in a way. They are algorithms which create graphs that resemble social networks and the like. –  Harry Stern Dec 9 '10 at 22:37
    
Here's another link: Fractal dimension on networks, specifically the cluster growing method might be more of the type of thing you're looking for? And just a small piece of (biased) advice, beware of the barabasi-alber and watts and strogratz model, I'm not sure they capture the power law degree distributed graphs you appear to be talking about. –  user4143 Dec 17 '10 at 4:25
add comment

1 Answer

Here is one example I stumbled upon.

Random walks have been used to identify clustering in networks in Rosvall and Bergstrom, Maps of random walks on complex networks reveal community structure. In particular, this technique was used by Eigenfactor (who publish journal rankings) to deduce clusterings in research communities.

share|improve this answer
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.