mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alessandro Negro <alenegr...@yahoo.it.INVALID>
Subject Re: Top-N recommendation with matrix factorization
Date Sat, 21 May 2016 11:07:26 GMT
Hi Mario,
In my opinion if you have a big matrix you can reduce its dimension using some SVD or PCA
method that allow you to reduce the size of the matrix. This should allow you to increase
performance reducing matrix factorisation time.

Mahout provides functions for both of them.

Br,
Alessandro 


> Il giorno 20 mag 2016, alle ore 12:59, Mario Levitin <mariolevitin@gmail.com> ha
scritto:
> 
> Hi,
> 
> If one is using a matrix factorization based method, in order to generate a
> top-N recommendation to a user, all the unknown ratings of that user needs
> to be predicted (so that highest predicted N items can be recommended). If
> we are talking about a site with millions of items this means that to make
> a top-N recommendation to a user, that user's rating on millions of items
> need to be predicted. This seems rather an inefficient way. I have two
> questions:
> 
> First one is general: do you know how this is can be done in a more
> efficient way, or how real large sites do this.
> Second, how can I do this efficiently with Mahout.
> 
> Thanks
> mario


Mime
View raw message