mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sean Owen <sro...@gmail.com>
Subject Re: Clarifying output from Taste demo
Date Fri, 20 Mar 2009 17:48:57 GMT
What do you mean by recommendations "given an item" -- I think the
answer is 'yes' with a little work. Do you mean

- recommended users that would like an item, given an item? Yes, this
would mean transposing the data model so that user IDs are item IDs
and vice versa, that's all

- items similar to a given item? doesn't need a recommender, just use
an ItemSimilarity. There is a convenience method on
ItemBasedRecommender that does this for you too

- recommendations from an "item-based recommender"? Sure just use
GenericItemBasedRecommender. Indeed, the point of using this
recommender is that you have some external, precomputed notion of item
similarity (which makes things fast, and injects more information).
But you can use a correlation-based similarity metric which is just
based on preference values too. It works, just doesn't really exploit
the point of item-based recommenders

On Fri, Mar 20, 2009 at 5:10 PM, Grant Ingersoll <gsingers@apache.org> wrote:
> Is there anyway in the demo to see item recommendations, given another item?
>  I think my quick perusal of the code showed only userID as input, but maybe
> I missed something.  BTW, that's item-item similarity, right?  And that is
> usually calculated offline, right?

Mime
View raw message