Hello everyone, I am Edward de Grijs.
While surfing on the internet I found this neat challenge of
a well defined problem, which I really like.
I started to work on this challenge at the end of Januari 2009.
I wished that I had known this sooner, but I didn't know that
this challenge existed.
With a goal of winning in my mind (who doesn't), I started to work
in my spare time on the SVD model, which seems to be the most
important model for recommender systems.
My thoughts were that if I wanted to win this competition, I must
have results which are better than the published ones.
For SVD I succeeded to make a fast and strong single model
that reaches a quiz score of 0.8767 with 500 features. (more features
will likely to improve this score even more, but that was not possible
on my computer because of memory requirements.)
To emphesize the speed, it is capable of reaching a score of 7.2%
improvement (qRMSE<0.8826) within 9 minutes when using all
four cores of a standard desktop Q6600 quadcore machine.
When a bunch of new users and/or movies are added to the dataset,
within 9 minutes the single model for all the new included users
and/or movies will be ready.
For a practical recommender system where the users and items
will change rapidly, this seems to be a very useful model, especially
because there is no need to try different adjustments.
It fully optimizes itself.
Stay tuned, more to come....
