2/23/06 Meeting

• Technical Report on Docu-scope
• Analysis of the Fedralist Papers
• Detecting Collaborations in the Federalist Papers

  12/1/05 Meeting

  Sorry about the sparse updates:) We discussed several ideas on future work, with an eye to staking out what papers people were interested in writing. The two ideas that got the most play were looking at multiple, overlapping networks, and looking at how to infer a network given data. In the latter case, a more useful, general form would be how to do an EM-type approach to take advantage of a partially observed network. When looking at these issues, the question of how to evaluate the fit of a network came up as another open topic. There was again a call for people to think about code they could write for Eric to run this summer assuming an internship at AT&T.

  11/3/05 Meeting

  Eric Vance gave a presentation on dynamic social networks. His animations are available here.

  10/27/05 Meeting

  Deepak Agarwal spoke about modeling social networks by call patterns. As a follow-up, he has sent a paper on how to detect subscription fraud.

  10/6/05 Meeting

  We went over the Nowicki and Snijders latent class paper. Some of the main topics for generalization/extension included:
  1. Handling the situation where the number of classes is unknown. This looked easy---put a prior of this, or try multiple values and assess fit.
  2. Allowing people to belong to more than one class. Edo and Stevie Fienberg have a paper on this, using a model in which a person can have fractional membership in multiple classes. This paper has been sent to the group and will be discussed at the next meeting. My own sense is that this is not yet an ideal solution; for many problems it seems better to let allow the sum of the membership fractions to exceed 1, or perhaps have binary values whose sum can exceed 1.
  3. There was discussion of allowing multiple sets of classes; religious affiliation gives one partition, political party another, education level a third. The Cartesian product of these can impact the number of parameters needed, unless strong (but perhaps reasonable) modeling assumptions are made. Eric and Edo pointed out that the Hoff, Raftery, and Handcock latent space paper may offer an alternative---a cluster analysis in the latent space could automatically determine co-class groups for multiple classes. But this probably loses some of the easy interpretability of the latent classes model.
  4. We considered applying both a latent class model and a latent space model simultaneously to the same data. The latent class could account for unobserved categories such as race or religion, whereas the latent space model could incorporate temperamental variables in friendship formation.
  5. We discussed building models for homophily and heterophily. The homophily is standard. Heterophily arises in opposite-sex attraction and also in hierarchical networks where people suck up to those above them on the ladder.
  6. We briefly discussed feedback mechanisms in social networks, a la the mechanisms used in network models for engineering and biology. The best examples so far are peer pressure or social disincentives to marry outside one's race/class. I hope we can revisit this as we move forward.

  9/29/05 Meeting

  Literature review. Michael Last will spoke about a paper by Adrian Raftery titled "Latent Space Approaches to Social Network Analysis". The approach proposed by Dr. Raftery is to model the probability of a link between two actors as a function of covariates and the distance between the actors. Given the observed links between actors, the distance needs to be estimated. Discussion focused on understanding the models proposed (distance and projection), especially the question of what did these distances correspond to in the real world? There was also some discussion of how to add dynamics to this model.

  9/22/05 Meeting

  1. Each member of the group introduced himself or herself, giving details of past and ongoing research on social networks.
  2. More detailed accounts were given by
     • Hugh Chipman, of research that he and colleagues are pursuing in NDHS settings
     • Steve Fienberg, of a sizable SN research group (numbering at least 12) and Carnegie Mellon
     • Alan Karr of research at NISS/NCSU/Duke under the MSBS grant to NISS, which is studying models for dynamics of social networks based on time-changing (latent or observable) characteristics.
  3. A number of research questions were identified (although how and by whom they will be pursued remains uncertain until they are understood more fully):
     • Scalability of models and inference procedures
     • Inference based on sampling of networks and other forms of incomplete observation
     • Networks with multiple relationships (two abstractions are multi-valued edges and multiple networks with overlapping sets of nodes)
     • The relative roles of observable and latent characteristics
  4. It was agreed to switch the primary meeting time to 1-2 PM Eastern time on Thursdays.

  Other useful papers and resources:

• CSSS Working Paper no. 46(1.5M bytes) [PDF(341K bytes )] "Model-Based Clustering for Social Networks" Mark S. Handcock , Adrian E. Raftery and Jeremy Tantrum April 2005., online here
• Work of Tom Snijders, online here
• Corinna Cortes, Daryl Pregibon and Chris Volinsky, Computational Methods for Dynamic Graph", Journal of Computational and Graphical Statistics. 12. 950-970, 2003. [computational methods specifically developed to look for unusual or fraudulent behaviour on a phone network]
• A processed version of the Enron data here
• "Introducing the Enron Corpus", by Klimt and Yang: pdf here
• "Exploration of Communication Networks from the Enron E-mail Corpus", Deisner and Carley: pdf here
• Resources for Variational Methods:
  1. An introduction to variational methods for graphical models. M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, and L. K. Saul. In M. I. Jordan (Ed.), Learning in Graphical Models, Cambridge: MIT Press, 1999. .. http://www.gatsby.ucl.ac.uk/~zoubin/papers/varintro.ps.gz
  2. M.J. Wainwright and M.I. Jordan. Graphical models, exponential families, and variational inference. Technical Report 649, Department of Statistics, University of California, Berkeley, 2003 .. http://www.cs.berkeley.edu/~jordan/papers/WaiJorVariational03.ps.gz

     A comprehensive reference. In particular, section 3 spells out the nature of the approximation of p(latent|params,data) with the variational distribution q(latent|variational_params(data)) in terms of Kullbach-Leibler divergence. Section 4 is on variational inference for mean parameters. Section 5 is on mean field theory in general.

  3. D.M. Blei, A. Ng, and M.I. Jordan. Latent Dirichlet allocation. Journal of Machine Learning Research, 3:993-1022, January 2003. .. http://www.cs.berkeley.edu/~blei/papers/blei03a.pdf

     A popular model, which makes a fully worked out example.

  4. M.J. Beal. Variational Algorithms for Approximate Bayesian Inference PhD. Thesis, Gatsby Computational Neuroscience Unit, University College London, 2003. (281 pages) .. http://www.cse.buffalo.edu/faculty/mbeal/papers/beal03.pdf

     More examples fully worked out can be found in Matthew's thesis.
     Ch 1: Introduction
     Ch 2: Variational Bayesian Theory
     Ch 3: Variational Bayesian Hidden Markov Models
     Ch 4: Variational Bayesian Mixture of Factor Analysers
     Ch 5: Variational Bayesian Linear Dynamical Systems
     Ch 6: Learning the structure of discrete-variable graphical models with hidden variables

  5. P. Xing, M.I Jordan and S. Russell. Graph partition strategies for generalized mean field inference. Uncertainty in Artificial Intelligence 20 (UAI2004), (eds. Meek and Halpern) AUAI Press, 602-610, 2004 .. http://www.cs.cmu.edu/~epxing/papers/Old_papers/xing_uai2004.pdfs

     Explores graph-cut methods to automatically partition a graphical model, in order to posit a variational distribution whose factors correspond to the partitions, to ultimately perform generalized mean field inference.