The newest recommender code runs on the new Scala R-like DSL. It is 
cooccurrence based and supports only LLR. LLR is used to downsample 
cooccurrences comparing all pairs of items. I’ve done fairly careful offline 
testing of all the similarity methods of Mahout’s hadoop and in-memory 
recommenders and LLR was a clear winner.

However if you have something new you want to try, look at the Scala 
SimilarityAnalysis class. For runtime efficiency it first calculates 
cooccurrences by performing [AA’] then calculating LLR on elements by row and 
downsampling in one step. You could look at some other similarity method for 
downsampling there. 

On Jan 16, 2015, at 12:44 AM, ARROYO MANCEBO David <[email protected]> 
wrote:

Any idea, Ted? :)

-----Mensaje original-----
De: Ted Dunning [mailto:[email protected]] 
Enviado el: jueves, 15 de enero de 2015 20:05
Para: [email protected]
Asunto: Re: Own recommender

The old Taste code is not the state of the art.  User-based recommenders built 
on that will be slow.



On Thu, Jan 15, 2015 at 7:10 AM, Juanjo Ramos <[email protected]> wrote:

> Hi David,
> You implement your custom algorithm and create your own class that 
> implements the UserSimilarity interface.
> 
> When you then instantiate your User-Based recommender, just pass your 
> custom class for the UserSimilarity parameter.
> 
> Best.
> 
> On Thu, Jan 15, 2015 at 1:11 PM, ARROYO MANCEBO David < 
> [email protected]> wrote:
> 
>> Hi folks,
>> How I can start to build my own recommender system in apache mahout 
>> with my personal algorithm? I need a custom UserSimilarity. Maybe a 
>> subclass from UserSimilarity like PearsonCorrelationSimilarity?
>> 
>> Thanks
>> Regards :)
>> 
> 

Reply via email to