Thanks all for the good info. Taste definitely sounds like a promising direction for us to go in for our recommendation service. I'm working through the installation of the GroupLens demo, but the mvn package step is failing with the output at http://paste.pocoo.org/show/110618/. Haven't looked into this yet, just thought I'd post to the list first with my progress. If anyone else uses IRC, I've created (and am currently the only one in) the #mahout channel on freenode. Hope to see some of you in there!
Josh On Wed, Apr 1, 2009 at 5:48 AM, Sean Owen <[email protected]> wrote: > Couple clarifications - > > The CF components are oriented to on-line, real-time use, though of course > one can trivially build a batch job out of that. That is what I did with > the > EC2 image that cranks out recommendations for all users. > > The CF component is also already parallelized as much as is practical. > There > are already Hadoop jobs for parallel, batch operation. > > Finally if you have some external notion of item similarity, like text > similarity between articles, you can and should include this info by > creating an ItemSimilarity with this knowledge. In that case you want to > use > an item-based recommender, since it is only in such a case that item-based > recommenders have a distinct advantage. > > On Apr 1, 2009 10:32 AM, "Otis Gospodnetic" <[email protected]> > wrote: > > > it's the former. Taste is still not parallelized, but other parts of > Mahout > are, and they make use of Hadoop. >
