[Topic-models] About collapsed Gibbs sampling
kchsjtu at yahoo.com
Thu Feb 19 21:03:47 EST 2009
I am reading the paper "Parameter estimation for text analysis" by Heinrich and have some questions about CGS (collapsed Gibbs sampling).
1. From the view point of Graphical models, is CGS equal to trimming those nodes that are integrated out and forming a new model on which all following derivation is based?
My question comes from p(w_i) in equation (76) in the paper, which I thought should be p(w_i) = p(w_i | z_(-i)). Why is w_i independent from z_(-i)?
2. Are there any rules about what kind of variables should be integrated in Gibbs sampling?
It seems that Dirichlet variables can be integrated out.
Thanks a lot and best regards,
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Topic-models