I'm just getting started on Mahout for a new project. I used Taste a few
years back, but things have changed a lot since then. So basically, I'll
be working on getting all the basic functionality I need first, and am
not really ready to take on such development right now.
I may look into persisting the transformation matrices if I need to, but
that's at least a few months away still. So if it's ready to use by the
time I need it, all the better ;-)
I'll be mostly working on integrating external user/content features
(demographics, etc.) to deal with cold start, and will rely as heavily
as possible on existing algorithms and implementations for the core CF
stuff.
Bye,
Jens
On 12/09/2011 03:20 PM, Sebastian Schelter wrote:
Yes, you describe it perfectly. I think the only reason this has not
been done yet is that the model computation is not very fast on Hadoop
because of its iterative nature.
Would you like to work on integrating the SVD recommenders?
--sebastian
On 09.12.2011 11:17, Jens Grivolla wrote:
On 12/08/2011 03:19 PM, Sebastian Schelter wrote:
[...]
A model for recommenders that use matrix factorization consists of the
user and item feature vectors. You can use a FilePersistenceStrategy
with any SVDRecommender to read and write these.
In the future we could also support loading the results of
ParallelALSFactorizationJob into an SVDRecommender.
I was actually looking for this. I guess this is the one case where
there actually is a "model", and calculating the factorization can be
costly.
I would expect that doing the "SVD" offline (e.g. on Hadoop) and then
providing online recommendations which only need a simple linear
projection is a pretty common use case, isn't it? You can even take new
user preferences into account in realtime (when projecting the user
vector into the feature space) with very little cost, and just update
the transformation matrices (which should be quite static) periodically.
Bye,
Jens
