I got my hands on the paper describing the system that is now the grand winner of the Netflix competition. The paper is written by

Yehuda Koren. Koren basically showed how to improve Netflix's recommendation algorithm by incorporating information about the changes in the ratings over time. I have only read a few beginning parts of the paper, but it looks very intriguing, and I am interested in knowing the details of the technique.

It is Yehuda Koren and not Kohen.

ReplyDeleteA single paper may describe only part of their system, which is in fact composed of over one hundred models, the results of which are blended to get the final result.

So it is not a single algorithm, and the time-related models are only a small part, with relatively minor contribution to the system.

You may want to refer to their other papers, which describe the more important models.

This comment has been removed by the author.

ReplyDeleteOops, sorry for the typo with the name. I know it is not a single algorithm, and should have made it explicit. I am currently looking at this paper, and will try to get hold of the other papers as well. Although, I am not sure how 'minor' the time-related model's contribution is, because it seems from the paper that the detailed modeling gave their system an edge. But, then again, I haven't looked at the many other models that comprise the system.

ReplyDelete