[Topic-models] Some questions about TOT
xuerui at cs.umass.edu
Wed Nov 26 15:10:34 EST 2008
1. Yes, you could learn LDA-like topic models without time and
analyze topics you learned with regard to time in a post-analysis
manner (see an example in Griffiths and Steyvers, Finding Scientific
Topics), but we want to study the dynamics of topic more
systematically, i.e., putting time into the modeling part. TOT and
DTM are two examples of this kind of effort, but the TOT model
utilizes time differently from state-transition based models (such as
DTM, cDTM, etc.). In all these models, topics discovered are
influenced by both word co-occurrence and the temporal information.
2. Other distributions could be used. The Beta distribution was used
based on several observations/intuitions: a) we don't want to
discretize time, and want a continuous distribution; 2) double-
bounded; 3) versatile enough to have different shapes/trends; and 4)
easy to parameterize, for inference convenience. If your time stamp
was discretized already, then you could relax a) if you prefer, for
instance. TOT model just shows that you can use time simply as a
continuous random variable. Similar ideas have been explored in other
papers as well, such as mixed membership model and supervised LDA.
Hope that the above explanation could help.
On Nov 24, 2008, at 1:58 AM, lykm wrote:
> When I read the paper "Topic Over Time", I find some question
> puzzled me. May someone can explain for me? Thanks a lot.
> 1 I think the meaning of the TOT is adding temporal information to
> the word as a tag, so we could drug the result from LDA and sort by
> temporal information, is it?
> 2 Why the author choose Beta distribution to generate temporal
> information? If not, how to choose the distribution, I mean what
> conditions should be satisfied by the distribution.
> I didn't find the answers from the paper, so anyone can help me ?
> Topic-models mailing list
> Topic-models at lists.cs.princeton.edu
More information about the Topic-models