Got it. Thanks, Sean! On Thu, Nov 17, 2011 at 11:42 AM, Sean Owen <[email protected]> wrote:
> Well I think you could fit it inside some of the user-user similarities, > yes. For a Pearson correlation, you could count important items twice or > something, yes. I wouldn't do that by literally adding more items to the > model as it creates other problems. It's possible; it may or may not have > the type and magnitude of effect you're looking for but easy enough to try. > > On Thu, Nov 17, 2011 at 6:25 PM, Jamey Wood <[email protected]> wrote: > > > I think that's certainly true for item-based recommenders (and item-item > > similarity). But isn't it a different story for user-user similarity? > In > > the example below, "novel1" and "novel1-copy" are indeed still separate > > items--but won't they be separate items that produce duplicative forces > > (and thus "weighting") in terms of the user-user similarity between user1 > > and user? > > > > I do realize that inflating the size of one's dataset in this way might > > lead to other problems. But setting that aside for now, I'd like to > > understand whether or not it would produce this kind of weighting effect > > for user-user similarities. > > > > Thanks, > > Jamey > > > > On Thu, Nov 17, 2011 at 10:59 AM, Sean Owen <[email protected]> wrote: > > > > > I don't think that would quite help, since novel1 and its copy are then > > > different items, and not somehow combining forces in the final > > calculation. > > > > > > On Thu, Nov 17, 2011 at 5:50 PM, Jamey Wood <[email protected]> > > wrote: > > > > > > > Thanks, Sean. We'll look into that. > > > > > > > > For user-based recommenders (or even just calculating > UserSimilarity), > > > > would it have the desired effect if we added multiple "virtual" > > > preference > > > > data points for the "real" items that we wished to more heavily > weight? > > > > For example, if our "real" preference data were: > > > > > > > > user1:novel1:3star > > > > user1:story1:4star > > > > user2:novel1:1star > > > > user2:story1:3star > > > > > > > > Would transforming it into this have the desired weighting effect (as > > > long > > > > as we filtered out the "copy" items in any actual recommendations)? > > > > > > > > user1:novel1:3star > > > > user1:novel1-copy1:3star > > > > user1:story1:4star > > > > user2:novel1:1star > > > > user2:novel1-copy1:1star > > > > user2:story1:3star > > > > > > > > The hope would be that "novel1" would now have twice the weighting as > > > > "story1" in determining the similarity of these two users. > > > > > > > > Thanks, > > > > Jamey > > > > > > > > On Thu, Nov 17, 2011 at 10:29 AM, Sean Owen <[email protected]> > wrote: > > > > > > > > > Not directly, but you could modify an item-based recommender to do > > so. > > > > > Where it uses an item-item similarity as a weight in a weighted > > > average, > > > > > you could modify the weight however you like depending on the types > > of > > > > the > > > > > two items. > > > > > > > > > > On Thu, Nov 17, 2011 at 5:16 PM, Jamey Wood <[email protected]> > > > > wrote: > > > > > > > > > > > Is there some way to weight particular preferences within Mahout? > > > For > > > > > > example, suppose you were creating some kind of literature > > > recommender > > > > > that > > > > > > uses a 5-star preference scale. If you wanted to give double the > > > > > weighting > > > > > > to preferences for novels versus preferences for short stories, > > what > > > > > would > > > > > > be the best way to do it? > > > > > > > > > > > > Thanks, > > > > > > Jamey > > > > > > > > > > > > > > > > > > > > >
