Posts Tagged 'preferential attachment'

Notes from ASONAM 2013

I’m at Asonam in Niagara Falls. I have noticed a few macro changes from the same conference last year in Istanbul:

  1. There is almost no interest in any form of clustering or community detection. I think that this is the result not of solving the problem but realising that it isn’t a well-formed problem for social networks (regular readers will be aware of my thoughts about this);
  2. There is growing awareness that preferential attachment does not generate networks that look very realistic, except when you look at them from a long way off with your eyes half closed (and some hope that models like forest fire might be close to usable);
  3. There has been a significant amount of progress in understanding the language use of tweets despite the obvious issues of dialect/patois, short length, mistyping and, often, lack of mental engagement when writing. I thought there was very little hope for this, so I’m delighted to be proved wrong. There are starting to be useful results learned from tweet corpora.

The mixture of attendees is less diverse than last year in Istanbul, not just geographically but by “home discipline”, which isĀ  a pity.

 

Advertisements

Problems at the heart of social network analysis

About a month ago, I was at the conference on Advances in Social Network Analysis and Modelling (ASONAM), a first for me. It’s a wide ranging conference with, for example, both sociologists and computer scientists presenting. Of course, “social networks” in this context means so-called online social networks.

I was surprised by the kind of presentations, and I came away thinking that there are two big problems at the heart of social network analysis that are unsolved and that are hardly being investigated, or even thought about:

  1. We can’t generate social networks artificially that look much like the real thing. Clearly, there’s been some progress here: preferential attachment gets much closer to the real thing than random graph models did; assortativity was another big step closer to reality; but it seems as if there must be at least one more big subtle process that we don’t understand about how real-world social networks get built. In other words, when humans form pairwise connections, there’s some aspect of that that we still don’t understand.
  2. There seems to be a universal assumption that an individual reveals, in his/her online social activity, a kind of homunculus of their total social activity — in other words, online behavior is a subset of real-world behavior. I only have to say this explicitly to show how fragile (maybe even foolhardy) such an assumption is. What people post on Facebook or tweet is a side-channel of their full behavioral spectrum — and an extremely odd side-channel as well. I haven’t seen any work in behavioral modelling that tries to understand how such a side-channel works and what it captures. But building models that make the subset assumption seems to me to be building castles on clouds.

There were some other, smaller assumptions that seemed to me to be accepted with too little thought as well. First, I don’t think that centrality measures have much to tell us in all but the smallest networks, because centrality implicitly assumes that a network has a single centre — but most networks have multiple centres whether or not they contain multiple communities (clusters); and anyway they mostly do.

Second, I think that the creation of a link is qualitatively a different kind of action to any subsequent use of that link — and so it is important to model them differently. This is quite tricky to do, it seems to me, but might repay the effort. Sometimes, the creation is a weak signal and its only the use of the link that makes it a noteworthy connection; friending someone is pointless unless there’s some actual interaction. On the other hand, marrying someone is a strong status-changing signal regardless of whether anything happens subsequently (for example, if nothing happens subsequently, immigration enforcement authorities may take an interest).


Advertisements