Hacker News Reader

I am curious why you used the user-closeness model.

Matrix factorisation won the Netflix prize: http://dl.acm.org/citation.cfm?id=1608614 . What made you go for user closeness?

View on HN

We've tested other models but collaborative filtering gave us the most "human"/natural results. Also, didn't Netflix toss the matrix factorization after the contest? IIRC they decided to keep their current algorithm.

Interestingly, my sense has been that Netflix's recommendation engine does a poorer job now than it did 10 years ago. I always assumed that it was because they used to use fairly straightforward collaborative filtering, and now they seem to be heavily focused on looking for stuff that's somehow cosmetically similar to other stuff I've watched.

So, like, instead of saying, "You liked Nosferatu? Well, other people who liked Nosferatu also liked Ran, so let's suggest that," it now goes, "Hey, that's a vampire movie! How about Blade?"

Reply on news.ycombinator.com