On Wed, Nov 2, 2011 at 9:40 AM, Ted Dunning <[email protected]> wrote:
> I have used synthetic data for testing SVD algorithms. It should be > reasonably easy to generate similar data of known shape for LDA. Not quite > as easy as generating matrices of known rank with known singular values, > but not much harder. > Oh, it's just as easy, actually: generate a "model" with a dirchlet prior, then start sampling your way (again using a dirichlet prior on your doc/topic dist) up to building a "corpus". Just hard for me to really know "how converged" is converged, as I can't use my euclidean intuition for distances. -jake
