[Topic-models] Questions regarding hLDA
blei at CS.Princeton.EDU
Sun May 18 12:16:25 EDT 2008
hi ben, rachit, and all,
> 4) A generalized hierarchical LDA with no fixed tree structure. Again
> the depth is L, and in this case I *really* don't see the point of
> this. But anyway, instead of a fixed number of children per node, use
> a CRP at each node controlling an indefinite number of children.
> Again to generate a document we first pick a path from root to leaf.
> But to do this, when we get to a node, we don't just randomly pick
> among the children. Instead, we consult the CRP, and it picks the
> appropriate child and returns the associated topic -- periodically
> creating new children/topics on the fly. Again, once you've got the
> path of length L, pick the mixing weights just as above and generate
> words as a mixture of L topics, again like above.
since the 2003 paper, we have extended this model to an infinite-depth
tree. the pre-print is here:
More information about the Topic-models