mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sean Owen <>
Subject Re: Combining user-based and item-based recommendation
Date Thu, 29 Oct 2009 13:03:46 GMT
Yes there are about five types of recommenders in the project
currently -- user-based, item-based, slope-one, clustering-based, and
one SVD-based implementation.

The difference between user-based and item-based isn't a question of
including stuff like demographic information. So what you're after is
not quite combining user-based and item-based approaches (though
separately, that is interesting).

What rating info could do is allow you to create a different, or
better, similarity metric rather than one based purely on rating data,
such as PearsonCorrelationSimilarity. Just write an implementation of
UserSimilarity that reads your demographic info, and for any two user
IDs, can return some notion of similarity based on that. Then you just
plug that into a user-based recommender.

Your implementation could be some kind of hybrid, too, between a
similarity computed from ratings and one from demographics. See how
averaging those two works. Plenty of possibilities here.

On Thu, Oct 29, 2009 at 12:18 PM, michal shmueli
<> wrote:
> Hi,
> I was playing with the examples and the demo, and to my best understanding,
> the recommendation is based on either "user-based" or "item-based". For
> example, the groupLens demo only looks on ratings but not on users data
> (i.e. demographics).
> I have 2 files, one for users and one for their ratings. Is there a way to
> combine them both and get better recommendation ?
> I'll appreciate if you can point me to such an example.
> thanks,
> Michal

View raw message