SVN? Or SVD?
With SVD, the easy way is to compute some number of singular values and if you reach some desired RMSE, then you quit. If not, you increase the number of singular values and start chugging again. I am not such a fan of this since my intuition for good RMSE isn't much better than my intuition for how many singular values are "enough". On Mon, Jun 28, 2010 at 5:59 PM, Tamas Jambor <[email protected]> wrote: > not sure how to do that with SVN. > > Tamas > > On 28/06/2010 23:10, Jake Mannix wrote: > >> Yep, this seems totally valid. In fact, even with the distributed SVD, >> adding the ability to stop after achieving a certain RMSE could be pretty >> helpful, so that people don't need to always pick what rank to ask for. >> >> Patches welcome! :) >> >> -jake >> >> On Mon, Jun 28, 2010 at 11:59 PM, Tamas Jambor<[email protected]> >> wrote: >> >> >> >>> depending what I for the initial step, with my setting initialSteps=60, I >>> can save around 50% with the MovieLens10m dataset. >>> >>> You can also justify this by saying that we are minimizing a function, so >>> that a fixed number of steps is not really the best way to achieve this >>> (e.g. depending on the slope it might arrive to the local minima sooner >>> or >>> later). >>> >>> Tamas >>> >>> >>> On 28/06/2010 22:31, Ted Dunning wrote: >>> >>> >>> >>>> How much speedup do you observe? >>>> >>>> On Mon, Jun 28, 2010 at 2:29 PM, Tamas Jambor<[email protected]> >>>> wrote: >>>> >>>> >>>> >>>> >>>> >>>>> Hi, >>>>> >>>>> I was looking at the SVD code, I am sure you are aware of this >>>>> modification, but it would really make things faster. The idea is that >>>>> you >>>>> set up a minimum RMSE improvement so it would stop training a >>>>> particular >>>>> feature if the algorithm reaches that threshold. I implemented it based >>>>> on >>>>> your code, it is just a slight change. >>>>> >>>>> Tamas >>>>> >>>>> >>>>> >>>>> >>>>> >>>> >>>> >>>> >>> >>> >> >> > >
