Is clustering still limited to mixture model hacks?
-
Tuesday, 1 May 2012 12:46 AM
Hi:
I'd like to use infer.net and I need to investigate change points and anomalies in temporal, well-correlated-high-dim, semi-big-data w/o labels. I was thinking of trying some recommended techniques in hiearchical subspace clustering as well as topic model variants. I was going to try the dynamic topic model of Blei-Lafferty, but it uses logistic normal in place of Dirichlet, so I can't even use LDA to hack an impl. Though, I suppose the Markov Topic Model of Wang, Thiesson, Meek, Blei could leverage LDA. Or for that matter, just leverage straight LDA into a subspace model. Now for more ordinary clustering, I was thinking of graph spectral and mean-shift, but I don't have a clue how to leverage infer.net? Of course, concurrently, I'll also play around with the usual dimension reduction techniques.
Assume I have strong motivation to try to have infer.net solve this problem.
I've already tried global LDA w/o success. I do know of several R packages that would really help at this point and perhaps the quickest route would be to solve the problem using R or other tools and then replicate on infer.net. But the point of this exercise is to only use infer.net. ;-)
What would you do? (I can give you any information you need, except the data.)
Thx -- Paul
