Thanks. yes, there are tricks but that would change SVD contract and amend the method, which is why I asked "without changing SVD contract", i.e. without changing what A, U, Sigma and V mean in SVD step.
But assuming we'd want to integrate PCA with the method, i guess yes special treatment of a mean side info could be devised. On Tue, Sep 6, 2011 at 12:35 PM, Ted Dunning <[email protected]> wrote: > Another option is to invent another kind of matrix that knows about an > offset. Then a special method for times may give the right performance. > > A third option is to do a little algebra on the PCA algorithm to propagate > the mean offset into the stochastic projection algorithm. > > On Tue, Sep 6, 2011 at 7:24 PM, Ted Dunning <[email protected]> wrote: > >> Sure. >> >> Do the subtraction after the B = Q'A step in the random projection! >> >> On Tue, Sep 6, 2011 at 7:16 PM, Dmitriy Lyubimov <[email protected]>wrote: >> >>> On Tue, Sep 6, 2011 at 12:11 PM, Ted Dunning <[email protected]> >>> wrote: >>> > Note that normally subtracting anything fills in sparse matrices. >>> >>> is there a way to cope with this without changing SVD contracts? >>> >> >> >
