This product needs to happen at recommendation time. And it is usually better to think of it in transposed form as the sum of a few columns selected by the non-zero elements of h. That way you retrieve just a few columns.
On Fri, Dec 4, 2009 at 10:59 AM, Sean Owen <[email protected]> wrote: > On Fri, Dec 4, 2009 at 6:04 PM, Jake Mannix <[email protected]> wrote: > > How would you do this? Take the rows of v_i cross v_i and add them > > up? Isn't that another MR job? > > No I suppose I meant take each row of A'A, one at a time, and dot with > h. That's computing the recommendation vector with only one row in > memory at a time. > -- Ted Dunning, CTO DeepDyve
