Hi Mac,

Collaborative Filtering algorithms only learn from interaction data
(known preferences) and are content agnostic, which means they don't
look at the actual content of the items.

This might sound awkward and counterintuitive at a first look but it
works really well when applied.

The relationship between "movie features" and ratings you use as an
example is already implicitly captured in the known ratings because
those have been given by people who might have used some criteria as a
basis for their rating.

I suggest you play a little with Mahout's CF implementations to see
whether you are content with the results.

If you just wanna know more about the theory behind Collaborative
Filtering and how implicit "features" can be dug out of the rating data,
I'd point you to "Matrix factorization techniques for Recommender
Systems" (http://www2.research.att.com/~volinsky/papers/ieeecomputer.pdf).

--sebastian


Am 01.10.2010 01:14, schrieb web service:
> I have got the group lens example working. Had a couple of doubts though -
> The dataset in grouplens has movieid, userid and the corresponding ratings.
> However a rating is meant to rate a movie but there are other things related
> to a movie to which the rating contributes.
> For example, the actors, directors, movie genre or may be the year of
> release etc.
> 
> So, is there any way to  capture this relationship and then generate
> recommendation.
> 
> Any suggestions, ideas about how to represent data or vector and then
> compute recommendations or how it is done usually etc. would be nice.
> 
> -Mac
> 

Reply via email to