Dear members,
I am doing a cluster analysis of some multivariate discrete data by fitting
a finite mixture of Multinomial distributions. I do this within a Bayesian
framework, and use a Dirichlet prior for each parameter vector of the (say,
L) Multinomial distributions. As it is standard with latent class models of
this sort, the (hidden) multinomial distributions are assumed to exhibit
statistical independence.
However, I would be interested in modeling the fact that the L multinomial
distributions determining the clusters are not independent. In particular, I
would like to come up with a model where each multinomial is itself
parameterized in terms of a *mixture* of the remaining L-1 components. For
each cluster, I would then be interested in estimating the L posterior
weights.
Can anybody pinpoint any relevant references? That would be very much
appreciated! Thank you all.
Cheers,
Frank
_________________________________________________________________
The new MSN 8: advanced junk mail protection and 2 months FREE*
http://join.msn.com/?page=features/junkmail
|