yes i was asking about lll branch in your repo. -d
On Tue, May 24, 2011 at 5:44 PM, Ted Dunning <[email protected]> wrote: > Heh? > > On Tue, May 24, 2011 at 5:12 PM, Dmitriy Lyubimov <[email protected]> wrote: > >> so you do in-memory latent factor computation? I think this is the >> same technique Koren implied for learning latent factors. >> > > Are you referring to the Log linear latent factor code that I have in my > mahout-525 github repo? > > Or SVD (L-2 latent factor computations)? > > Or LDA (multinomial latent factor computation)? > > Or Dirichlet process clustering? > > Or ...? > > >> However, i never understood why this factorization must come up with r >> best factors. I understand incremental SVD approach >> (essentially the same thing except learning factors iteratively >> guarantees we capture the best ones) but if we do it all in parallel, >> does it create any good in your trials? >> > > I don't understand the question. > > Are you asking whether the random projection code finds the best (largest) > singular > values and corresponding vectors? If so, the answer is yes, it does with > high probability > of low error. > > But this doesn't sound like LLL stuff. > > >> also i thought that cold start problem is helped by the fact that we >> learn weights first so they always give independent best >> approximation, and then user-item interactions reveal specific about >> user and item. However, if we learn them all at the same time, it does >> not seem obvious to me that we'd be learning best approximation when >> latent factors are unkown >> (new users). Also, in that implementation i can't see side info >> training at all -- is it there? >> > > This sounds like LLL again. > > In LLL, the optimization of side information coefficients and the > optimization of the user-item interactions are separately convex, but not > jointly convex. This means that you pretty much have to proceed by > optimizing one, then the other, then the first and so on. > > So I don't see how we are learning them all at the same time. > > You are correct, I think that having lost joint convexity that we don't have > strong convergence guarantees, but practically speaking, we get > convergence. > > Regarding the side information, I thought it was there but may have made a > mistake. > > Sorry for being dense about your message. >
