The new Mahout recommender code doesn’t attempt to predict ratings at all. It 
tries to predict which items are of most interest to users. Ratings are a bit 
of a red herring.

If you want to learn about recommenders the old Mahout in Action text is out of 
date so I wouldn’t follow that as a learning tool.

Again I’d point to the references at the top of this page, including an easy to 
read book by Ted.
http://mahout.apache.org/users/recommender/intro-cooccurrence-spark.html

This describes a recommender that can ingest most of a user’s clickstream, 
correlate user actions to some canonical action that you _do_ want to recommend 
(like watching a video) then makes the best ranking of recommended items. It’s 
multimodal because it can take actions like watching a video, rating a video, 
searching, being at some location, browsing categories, and turn the data into 
a prediction about what video the user is likely to want to watch. 

On Feb 16, 2015, at 1:32 AM, Eugenio Tacchini <[email protected]> 
wrote:

Hello,
I was just referring to prediction in its classical meaning: you have a set
of ratings (users rate items) and you want to predict one of the missing
ratings e.g. P(User A, Item 23).

Eugenio

2015-02-13 21:04 GMT+01:00 Pat Ferrel <[email protected]>:

> spark-rowsimilarity will give you a list of similar users (rows in the
> interaction matrix) using LLR with several downsampling options. This works
> with rows for input but you can input elements with a little custom code to
> get exactly the same result.
> 
> Let me understand the second part of your question. The recs query is
> (user id, item id)? So you want both to contribute to the recommendations?
> This is different than a typical “other people who like this also liked
> these” type rec set, which is non-personal—the same for every user.
> 
> If you are asking for something like recs on a product page using the item
> being viewed as context and the user’s preference history too—the
> multimodal recommender can do that. But please explain before I go into a
> long reply.
> 
> On Feb 13, 2015, at 9:53 AM, Ted Dunning <[email protected]> wrote:
> 
> On Fri, Feb 13, 2015 at 9:37 AM, Eugenio Tacchini <
> [email protected]> wrote:
> 
>> If I need to use a classical user-based technique, however, the only
>> alternative is the Taste-oriented code, am I right?
>> 
> 
> Right.
> 
> 
>> Still, I can't see how
>> to perform a prediction for a a user/item couple, is there a class for
>> that?
>> 
> 
> Not directly, but I think that you cna cobble something simple together.
> 
> 

Reply via email to