Yes. And this last operation is only separate because it so often needs access to row sums and overall totals. It could be inserted into the first MR step at the cost of clarity.
On Wed, Sep 9, 2009 at 8:48 PM, Sean Owen <[email protected]> wrote: > Yes -- that second mapreduce is the real question! is there such a > parallelizable black box that transforms the matrix multiply result > into recommendations? I'm guessing there is indeed a matrix that makes > the result into useful similarity weights, then this can be finished > off by multiplying by the user rating vector. > -- Ted Dunning, CTO DeepDyve
