JustHackIt

I am interested in applying vectoral semantics topic models (e.g. LDA) to threaded message boards (e.g. most blogs, usenet, reddit, digg, HN etc.)

Things I would like to try:

(1) classify users based on the topics of their comments and or the topics of their parent or child messages.

(2) see if any of the topics can be used to make meaningful classifications of user tastes and values (this has already been done using decision trees to estimate the demographic identity of bloggers)

(3) see if there is any "phylogenetic structure" in the topics of a message thread. Is it possible to automatically identify which topics are most likely to provoke indignant responses or which topics are most likely to provoke puns or one liners?
Models like the Kingman coalescent seem applicable.

My strategy is to approach this as a research project. I have a decent amount of experience in C++ and python for scientific / numerical applications, but my CS background is weak. I'd like to talk to a more hardcore CS person with AI / ML experience to give my work some better feedback. I would prefer to work with python or Ocaml.

This is very much a background / side project.