On Tue, Jul 19, 2011 at 1:07 AM, Sebastian Schelter <[email protected]> wrote:
> On 19.07.2011 01:32, Ted Dunning wrote: > >> Well... it makes it uncomputable in explicit form. Sometimes there are >> implicit forms for the matrix that keeps the size in bounds. For >> instance, >> limited rank decompositions (aka truncated singular value decompositions) >> can be represented by storing two skinny matrices and a diagonal that >> don't >> take much more memory/disk than the original data. >> > > Aren't the user/item feature matrices already a form of this? > Yes. > I think the basic question was how to compute recommendations for a > particular user. You could just predict his preference for all items he has > not yet seen by multiplying the item features matrix with his user feature > vector to get the estimated preferences but you cannot to this for all users > at once, right? > If you trim the recommendations to the top N, you probably can do this computation. > So the question would be how to find the initial "candidate items" from the > item feature matrix. > Precisely put. You can do this ahead of time and save vats of work or just do the multiplication which will take a while to do, but you only need to keep one dense vector at a time.
