[Topic-models] HLDA, is Stick Breaking Prior Necessary?
ccrraappy at gmail.com
Wed Mar 24 22:22:20 EDT 2010
I tried an implementation without stick breaking prior. Like what you
said, the stop words do not appear on top. Thanks alot!
As for my id ccrraappyy , coz I figured out that the world is a crappy
place. So I wanted to use nicks like crappy but it is always taken by
someone else. I am not creative enough to spend effort and time to
figure out some new nicks so i just double the letters in it. But the
older systems have a limit on user id length, so rather than
ccrraappppyy it became ccrraappy.
Later, someone told me crappy really means shitty and it had never
occur to me before that this nick is so obnoxious. LOL.
On Thu, Mar 25, 2010 at 2:55 AM, David Blei <david.blei at gmail.com> wrote:
> dear jake,
> first, it is necessary for considering infinite length trees.
> second, in practice, it helps encourage higher level topics to be more
> "abstract." otherwise, the posterior has the flexibility to associate
> a higher level topic with (non-abstract) words from a document, and to
> simply not associate other documents with that topic. (i've seen this
> happen.) you are right that perhaps higher posterior probability is
> obtained with the "right" decomposition, but the sampler is easily
> stuck in a variety of local minima.
> PS. i have to ask: why the email address "ccrraappyy at gmail.com"?
> On Wed, Mar 24, 2010 at 7:27 AM, Jake Sully <ccrraappy at gmail.com> wrote:
>> Hi List,
>> I am back with more questions :)
>> I have been wondering, is the stick breaking prior really necessary?
>> Since the documents are already sharing nodes at the top of the tree,
>> do we still need to add more statistical pressure to force the words
>> to be near the top. I am just wondering whether anyone have tried.
>> Topic-models mailing list
>> Topic-models at lists.cs.princeton.edu
More information about the Topic-models