[Topic-models] fast gibbs sampling

Gregor Heinrich gregor at arbylon.net
Wed Mar 2 21:12:15 EST 2011


hi chi yuan -- a denominator of akj would not be dependent on the sampling 
dimension k.

gregor

On 3/3/11 3:01 AM, chi yuan wrote:
> Hi people,
>
> I read about  Fast Collapsed Gibbs Sampling For Latent Dirichlet Allocation by
> Ian Porteous. However, one thing that confuses me is that for his equation for
> collapsed gibbs sampling -
>
>
> p(zij = k|z¬ij , x, α, β) = 1/Z * akj  * bwk
>
>   whereby akj is Nkj + alpha , Nkj is number of times word in document k has been
> assigned to topic j,
>
> bwk = (Nwk/ Beta)/(Nk + W * Beta), Nwk is the number of times word w has been
> assigned to topic k
>
>
> My question is:
>
> Why does he normalized the akj part of the equation like what he did with bwk? I
> was referring this to the collapsed gibbs sampling equation I saw in 'Finding
> Scientific Topics' and realize this difference.
>
>
> Chi Yuan
>
>
> _______________________________________________
> Topic-models mailing list
> Topic-models at lists.cs.princeton.edu
> https://lists.cs.princeton.edu/mailman/listinfo/topic-models


More information about the Topic-models mailing list