Hello Kumar,
Unfortunately the paper is the best documentation available for the ALS
algorithm (together with the unit tests) and a good choice of parameters
is to be found by experimentation.
There is also a script available that applies the factorization to the
movielens dataset: mahout-examples/bin/factorize-movielens-1M.sh
I suggest reading the article "Matrix Factorization Techniques for
Recommender Systems" by Yehuda Koren that offers a nice to read
introduction to matrix factorization in CF.
http://research.yahoo.com/pub/2859
There is no video or reference tutorial available regarding ALS or
RecommenderJob but using the latter should be pretty straight forward.
Feel free to ask your questions here.
--sebastian
On 20.06.2011 21:50, Kumar Kandasami wrote:
Hello Sebastian:
I was going over the ALS-WR paper (by Yunhong Zhou, Dennis Wilkinson,
Robert Schreiber and Rong Pan- HP Labs on Netflix dataset) this weekend,
and I am still trying to understand the algorithm.
I am currently working on running item based recommender on the Wikipedia
link dataset (boolean preferences) on EC2 clusters.I am interested in
testing the ALS recommender, however, at this point I have no clear
understanding of what user/item features mean, and even determining the
iterations as well as numoffeatures attribute upfront.
Is there any documentation or overview on the usage of the Mahout ALS
implementation ?
Additionally, it will save us lot of time, if you could forward any
reference tutorial or video presentation links (similar to Item-similarity
Job on vimeo) on RecommenderJob or ALS
Kumar _/|\_
www.saisk.com
[email protected]
"making a profound difference with knowledge and creativity..."
On Thu, Jun 16, 2011 at 11:48 AM, Sebastian Schelter<[email protected]> wrote:
Hello Kumar,
Check the Mahout JIRA for features planned for 0.6 at
https://issues.apache.org/**jira/browse/MAHOUT<https://issues.apache.org/jira/browse/MAHOUT>
It would be great if you could test the distributed ALS recommender that
uses matrix factorization. If you wanna dive into that I'm sure we'd find a
lot of things you could improve.
Check it's original jira issue as a starting point:
https://issues.apache.org/**jira/browse/MAHOUT-542<https://issues.apache.org/jira/browse/MAHOUT-542>
If you want something small to work, you can check
https://issues.apache.org/**jira/browse/MAHOUT-609<https://issues.apache.org/jira/browse/MAHOUT-609>
Does that match what you expected? If you have any ideas yourself, feel
free to share them with us.
--sebastian
On 16.06.2011 18:06, Kumar Kandasami wrote:
Hi !
Could anyone point to a link that outlines the features expected in the
mahout 0.6 release ? Specifically distributed recommendation engines.
Also, I am currently using/working on the distributed recommendation
engines
on EC2 clusters - is there a way that I could contribute any code that
would
be in the 0.6 or future road map.
Thank you.
Kumar _/|\_
www.saisk.com
