I solved my problems using a hint from a recent post: calling the RandomUtils.useTestSeed() before building the recommender returned consistent results.
On Sat, Mar 23, 2013 at 4:33 PM, Nadia A Najjar <[email protected]>wrote: > I'm using the 100k ML data. I need to generate predictions using SVD++. In > testing my setup I randomly chose a user and an unrated item to for that > user then estimate preference. > for example, estimating preference for user 1 and item 274 with a seeded > random generator, one run returns 3.4404 and another 3.3961. This was using > 19 factors (number of genres in ML) and 60 iterations. > Even though the difference is small for my research I would like to > eliminate this difference. > For seeding the factorizer I seeded the random generator in the > prepareTraining () method in SVDPlusPlusFactorizer class. > > > > > From: [email protected] > > Date: Sat, 23 Mar 2013 14:16:10 +0100 > > Subject: Re: Persistent SVD predictions > > To: [email protected] > > > > For relatively small amounts of data, the effect of randomness may be > > large. For larger amounts of data, the effects should be small. > > > > On Sat, Mar 23, 2013 at 1:35 PM, Sebastian Schelter < > [email protected] > > > wrote: > > > > > Hi Nadia, > > > > > > This should be the case if you fix the seed for the random number > > > generator. What do you mean by the "results differed"? How did you > > > choose the single item that you looked at? How did you choose the > number > > > of iterations? > > > > > > Best, > > > Sebastian > > > > > > > > > > > > > > > On 23.03.2013 13:10, Nadia A Najjar wrote: > > > > Hi, > > > > to rephrase my question: > > > > Does the SVD++ factorization produce the same prediction value for a > > > given item every time the SVD recommender is run? > > > > > > > > I'd appreciate any insight on this! > > > > > > > > Thanks! > > > > > > > > > > > >> From: [email protected] > > > >> To: [email protected] > > > >> Subject: RE: Persistent SVD predictions > > > >> Date: Wed, 20 Mar 2013 01:32:35 -0400 > > > >> > > > >> Hi,This is when comparing a single prediction for an item. I seeded > the > > > initial model training. That was the only place I found where a random > > > number was used. > > > >> > > > >>> Date: Tue, 19 Mar 2013 19:34:16 +0100 > > > >>> From: [email protected] > > > >>> To: [email protected] > > > >>> Subject: Re: Persistent SVD predictions > > > >>> > > > >>> Hi Nadia, > > > >>> > > > >>> how did you try to seed the factorization? How do you measure your > > > result? > > > >>> > > > >>> > > > >>> On 19.03.2013 19:04, Nadia A Najjar wrote: > > > >>>> Is there a way to get the same preference prediction result for an > > > item using SVD recommender with SVD++ factorization?I tried seeding the > > > factorization but the results still differed! > > > >>>> Thanks! > > > >>>> > > > >>> > > > >> > > > > > > > > > > > > > > > >
