*I think* there's some space to handle regression hierarchies and generalize combining methods like here http://www.ideal.ece.utexas.edu/seminar/LatentFactorModels.pdf via mere architectural enhancements to solve those problems but it's all just an intuition. In particular, I understand this yahoo work as some sort of a hybrid of factorization, regression on content data and regression on side data, and conceivably one might generalize framework enough to plug method that is doing factorization from those doing regression. Somehow. E.g. it probably doesn't matter if we want to handle factorization work using ALS or incremental SVD. We could build an overarching model that would take that as a parameter for an algorithm class. Perhaps.
Or perhaps there's not enough space for generalizing this. On Tue, Mar 22, 2011 at 3:06 PM, Robin Anil <[email protected]> wrote: > Well thats the implementation detail of the system, there are various > strategies, driven by various aspects of the product. Not generic like the > algorithms in Mahout. I dont think its something Mahout should handle > anyway. > It will be great if Foursquare folks on the list can share > stories/issues/bugs or even patches they made in getting Mahout jobs to run. > Should help us polish Mahout for a v1.0 milestone. > > On Wed, Mar 23, 2011 at 3:29 AM, Dmitriy Lyubimov <[email protected]> wrote: >> >> Hm. they say they had a cold start problem too that Mahout doesn't handle >> yet. >> >> On Tue, Mar 22, 2011 at 2:34 PM, Robin Anil <[email protected]> wrote: >> > >> > http://engineering.foursquare.com/2011/03/22/building-a-recommendation-engine-foursquare-style/ >> > > >
