My NefFlix ratings: how do they influence "Movies You'll Love"?

[Argelius]

I'm interested in how the algorithm for NetFlix ratings<=>Recommendations works.

I like perusing the "Movies You'll Love" section on NetFlix to find movies I might otherwise never rent.

I'm curious, if anyone can possibly know these things, how my ratings influence what I see there. For instance, say I really like Sci-Fi movies, but I rent one that is awful and give it a one-star rating (even though it was the kind of movie I like, but just a bad example of it. Does that then reduce the likelihood that Sci-Fi movies will be recommended to me? I'm guessing it's more complicated than that...

[Mort in Detroit]

I think it's far more complicated than that. I don't think they really spell it out very well. Proprietary reasons?

I've been a member for over five years. Their predicted ratings, to me, are fantastic. I wouldn't say the same about Blockbuster or Yahoo. I think where some get into trouble is that they don't rate on how much they enjoyed the film--they rate the film on how good they think it was objectively. I've rated films that I knew were excellent films 2 or 3 stars because I either didn't really care for them or was only mildly glad I saw them. I think that's what keeps predictions a little more accurate--and I have had my eyes opened to and greatly enjoyed films of every genre from many nations that I'd have never otherwise seen if it weren't for Netflix recommendations. I'd never seen films from Iceland or with John Wayne (who I'd always thought was a talentless git) among many other classifications and I really enjoyed. When I gave the John Wayne movies a high rating, my predicted rating for most other John Wayne films remained low. (One of many examples.) That leads me to believe their system is very sophisticated.

[Adrian]

If you want to know more about how it works, there is a lot of detail in the Netflix Prize site.

The best way to think about it is that given a few billion movie ratings we can take the set of movies that you have rated, and say that there are some people who have rated the same movies the same way, and those people have also rated movies that you have not rated, so we can use their ratings to predict what you will like. The algorithm is a mathematical pattern match, it doesn't actually know anything about movies as such, which is why some of the recommendations are hard to explain. The big trick is to make it scale to millions of customers and make it accurate enough to be useful.

Adrian from Netflix

[Knaldskalle]

Actually, there was a very interesting piece on Netflix ratings and the Netflix Prize in the NY Times the other day. I gave Napoleon Dynamite 2 stars.