[Topic-models] Topic-models Digest, Vol 49, Issue 7
wray.buntine at nicta.com.au
Thu Aug 19 19:26:09 EDT 2010
Gregor Heinrich wrote:
> Dear list experts --
> I was wondering whether someone has used or published a Monte-Carlo
> approach for the parameter of the exchangeable Dirichlet distribution
> (LDA's scalar alpha).
I'm pretty sure the distribution of these is log-concave so you can use
adaptive rejection sampling methods.
We take an entirely different approach, however for this when a full
Dirichlet is used, so for non-symmetric alpha. You can replace the
Dirichlet by a Dirichlet Process or Pitman-Yor Process and then use an
efficient collapsed Gibbs
sampler. Lan Du (ANU&NICTA) gets great results using this
(at another spot on the LDA) and its published in ECML-PKDD 2010, this year.
This collapsed Gibbs sampler is about 20 times faster than previous methods
for sampling with a DP or PYP.
Statistical Machine Learning
NICTA | Locked Bag 8001 | Canberra ACT 2601
T +61 2 6267 6323 | F +61 2 6267 6230
www.nicta.com.au | wray.buntine at nicta.com.au
>From imagination to impact.
The information in this e-mail may be confidential and subject to legal professional privilege and/or copyright. National ICT Australia Limited accepts no liability for any damage caused by this email or its attachments.
More information about the Topic-models