On Tue, Dec 8, 2009 at 6:49 PM, Zhengguo 'Mike' SUN <[email protected]>wrote:
> I appreciate if you could provide some pointers. > Of course, we'd also appreciate it if you could contribute that code here to Mahout, as well, if possible. It'd be great to have that here for everyone to use, as I'm sure that having the choice of doing NMF as well as SVD or LSI or pLSI would very helpful for the rest of the community! Patches welcome! Even if your first impls don't do any of the tricks we're describing, if you contribute it, we can help you improve it. -jake > > > ________________________________ > From: Jake Mannix <[email protected]> > To: [email protected] > Cc: [email protected] > Sent: Tue, December 8, 2009 9:08:19 PM > Subject: Re: Fwd: A MapReduce Algorithm for Matrix Multiplication > > On Tue, Dec 8, 2009 at 5:56 PM, Ted Dunning <[email protected]> wrote: > > > NMF should be amenable to the stochastic decomposition trick. That > reduces > > the problem to a much smaller factorization that you could plausibly do > > using sequential techniques. Jake Mannix is working on getting > > factorizations going, but I don't know if he has gotten to NMF. > > > > I'm not currently working on NMF, but the stochastic decomposition trick > will be > in there soon, which should allow all this pretty easily. > > Although... if you start with a positive matrix, you may want a specialized > random > projector which preserves positivity for this kind of thing. But I'm not > sure, I > haven't looked too closely at what happens when you try to do this trick on > NMF. > > -jake > > > > >
