You can take a look at Buntine and Jakulin, Discrete Components Analysis. http://cosco.hiit.fi/Articles/buntineBohinj.pdf
Note that the optimizations are nice, but the basic Gibbs sampler is actually pretty simple. My first implementation was about 20 lines of R and it wasn¹t actually so slow for training. With a decent cluster, the simpler algorithms may actually be better. (btw, if I include the full reference, I get kicked out with high spam score. Apparently names with initials or something is bad). On 4/15/08 7:22 PM, "Ian Holsman" <[EMAIL PROTECTED]> wrote: > Ted Dunning wrote: >> We should consider changing algorithms. >> >> MDCA is a good candidate. So would be nested Dirchlet processes. Neither >> of these is necessarily all that much more difficult to implement than PLSI >> and both should give better results. >> >> > Hi Ted. > can you give me a pointer to something that describes MDCA ? all the > things google finds is behind a paywall. > > regards > Ian >
