On Thu, Jan 10, 2013 at 8:32 PM, James Frohnhofer <fij...@gmail.com> wrote:
> > .... I'm not sure if I can come up with test data which exercises what I > saw, 2) > I can't vouch for the correctness of the change made. However, it does > behave the way I expect now. > Can you provide a test case that illustrates your needs? One thing that's bugging me though is that if I'm the first person to come > across this, is there another part of the library everyone else is using > for sparse linear regressions that I don't know about? > Single machine sparse regressions are relatively unused in Mahout. I know that when I need to do this, I typically would use an in-memory stochastic SVD. Even that use would be quite rare. I added LSMR some time ago anticipating a need for reverse engineering the work done by the feature hashing vector encoders. As it turns out, least squares is probably not quite the right solution for this and if this becomes important, we will probably need an least-L1 solver instead. That need is not turning out that large, however, so there has been very little use of LSMR.
