Unfortunately my user-user similarity information is quite limited too, it probably covers only 50% of the user base.
I was actually just reading TreeClusteringRecommender and TreeClusteringRecommender2, I was thinking to get a small number of clusters then pick top items from each cluster to create some diversity. I might do a A/B testing with popularity versus above method to see if there is any significant difference for the user. On Tue, Apr 20, 2010 at 4:58 PM, Sean Owen <sro...@gmail.com> wrote: > Ah well, if you have a priori user-user similarity, you can do > user-based recommendation for your new user even with no user-item > links for him/her. As long as you know user-user similarities you're > OK. > > Ted's suggestion is essentially a variant of this. You could use > TreeClusteringRecommender to do what he says. > > Your second question is a bit different from a question of > recommendation. Perhaps you would base such a list on *recent* > popularity? or recent positive change in popularity? You could > populate it with things that used to be popular? > > On Tue, Apr 20, 2010 at 9:25 PM, Tolga Oral <tolga.o...@gmail.com> wrote: > > PlusAnonymousUserDataModel will work once the user clicks on couple items > on > > the site, however still doesnt solve the dead-start problem. We are > creating > > user similarities based on different attributes and use the similarities > to > > recommend items (doesn't solve all cases though) > > > > However I am still interested in figuring out the most popular items with > > some diversity (otherwise new "interesting/good" items have no chance of > > ever getting in recommendations) ? Any ideas how we can do this in > mahout? > > >
