|
Hi all, I would like to seek help if someone has experience in estimating the K in LDA? I watch David Blei video on videolectures.net as he mentioned that it is achievable but was not described in the PPT or video. I google around and found that Hierachical LDA is one of doing it but not much code was found, can anyone provide some more detail information, thanks. Regards, Andy. |
|
I think Indian Buffet Processes do what you want, see cocosci.berkeley.edu/tom/papers/ibptr.pdf. Not really, the correct nonparametric extension to LDA is given by the Hierarchical Dirichlet Process. For a practical algorithm see http://www.cs.princeton.edu/~chongw/papers/WangPaisleyBlei2011.pdf
(Jan 07 '12 at 12:16)
Alexandre Passos ♦
|
|
Aside from nonparametric methods, if you can estimate the held-out likelihood after fitting your model you can search for the value of k which maximizes this held-out likelihood. For computing that likelihood see, for example, Wallach et al Evaluation methods for topic models. |
|
K is the number of topics, right? Non-parametric methods are typically used for this. Look at the hierarchical Dirichlet process and extensions there of. |
