Netflix Prize May Have Been Achieved

MadAnalyst writes "The long-running $1,000,000 competition to improve on the Netflix Cinematch recommendation system by 10% (in terms of the RMSE) may have finally been won. Recent results show a 10.05% improvement from the team called BellKor's Pragmatic Chaos, a merger between some of the teams who were getting close to the contest's goal. We've discussed this competition in the past."

Comments

1 Million split 7 ways

[basementman]

Let's see, $1,000,000 split 7 ways gives us $142,857.14 each. Let's say taxes take half, now you are down to only $71,428.57 each. Unless one of them kills all of their partners like in The Dark Knight that ain't much of a prize.

Re:1 Million split 7 ways

[quanticle]

Well, just like the Ansari X Prize didn't cover the costs of developing and launching a suborbital rocket, the Netflix Prize isn't really meant to be a large enough prize to fully fund the development of a new recommendation algorithm. The purpose of the prize is to stimulate interest and get people started. The real reward will come when they turn their algorithm into commercialized software - the rewards from making such a thing applicable outside of Netflix could be large indeed.

Re:1 Million split 7 ways

[neokushan]

Pretty sure having it on their CV means they can effectively write their own pay cheque in terms of job opportunities.

Well done!

[Slurpee]

Well done Bellkor.

But now the real race begins.

Now that the 10% barrier has been reached, people have 30 days to submit their final results. At the end of the 30 days, whoever has the best result wins.

This is going to be a great month!

Re:Do they keep the prize money?

[lee1026]

AT&T have committed to giving all money to charity. The person at yahoo developed his entry while working at AT&T, so I will be surprised if yahoo gets any of it.

they were able to get the extra 0.5% over the top

[circletimessquare]

by simply ignoring data from anyone who ever rented SuperBabies: Baby Geniuses 2, Gigli, From Justin to Kelly, Disaster Movie, any movie by Uwe Boll and any movie starring Paris Hilton

suddenly, everything made sense

Re:No info about the Netflix prize

[Reikk]

Background: The Netflix Prize is an ongoing open competition for the best collaborative filtering algorithm that predicts user ratings for films, based on previous ratings. The competition is held by Netflix, an online DVD-rental service, and is opened for anyone (with some exceptions). The grand prize of $1,000,000 is reserved for the entry which bests Netflix's own algorithm for predicting ratings by 10%.

most of them did it

[circletimessquare]

for simple intellectual satisfaction, like a giant puzzle or a game of chess

money is not the motivation for everything in this world

Re:No info about the Netflix prize

[MrMista_B]

What, you didn't even read the /summary/?

I know, this is Slashdot, but 'some basic info about it' is /right there/.

Interesting

[coaxial]

I published a paper using Netflix data. (Yeah, that group.)

It's certainly cool that they beat the 10% improvement, and it's a hell of a deal for Netflix, since it would have cost them more than a prize money paid out to hire the researchers, the interesting thing is whether or not this really advances the the field of recommendation systems.

The initial work definitely did, but I wonder how much of the quest for the 10% threshold moved the science, as opposed to just tweaking an application. Recommender systems still don't bring up rare items, and they still have problems with diversity. None of the Netflix Prize work address any of these problems.

Still, I look forward to their paper.

Re:Interesting

[Trepidity]

Trying to recommend unpopular movies is problematic. Is the computer program going to be able to discern under-rated (Glengarry Glen Ross) or just crap (Ishtar)

That is indeed an interesting question, and I think it's what the grandparent meant when he pointed out Netflix's contest didn't really address it. The performance measure Netflix used was root-mean squared error, so every prediction counts equally in determining your error. Since the vast majority of predictions in the data set are for frequently-watched films, effectively the prize was focused primarily on optimizing the common case: correctly predict whether someone will like or not like one of the very popular films. Of course, getting the unpopular films right too helps, but all else being equal, it's better to make even tiny improvements to your predictions of films that appear tons of times in the data set, than to make considerable improvements to less popular films' predictions, because the importance of getting a prediction right is in effect weighted by the film's popularity.

You could look at error from a movie-centric perspective, though, asking something like, "how good are your recommender algorithm's predictions for the average film?" That causes you to focus on different things, if an error of 1 star on Obscure Film predictions and an error of 1 star on Titanic predictions count the same.

Re:No info about the Netflix prize

[bogjobber]

If the first sentence didn't explain it enough, perhaps you could RTFA.

i was joking, however

[circletimessquare]

from the excellent nyt article about the competition in november:

http://science.slashdot.org/article.pl?sid=08/11/22/0526216

it isn't bad movies that are the problem, taste in bad movies can still be uniform

the real problem is extremely controversial movies, most notably Napoleon Dynamite

http://www.imdb.com/title/tt0374900/

not controversial in terms of dealing with abortion or gun control, but controversial in terms of some people really found the movie totally stupid, while some people really found the movie to be really funny

movies like napolean dynamite are genre edge conditions, and people who apparently agree on everything else about movies in general encounter movies like this one and suddenly dramatically differ on their opinion of it, in completely unpredictable ways

Re:i was joking, however

[bill_mcgonigle]

Yeah, all the recommendation systems where I've bought or rented movies, and most of my friends all said I needed to see 'Fight Club', so I did, and ... meh.

Consider this list of movies I've bought/rated highly:

  12 Monkeys
  V for Vendetta
  Lost in Translation
  Donnie Darko
  A Beautiful Mind
  Dogma

I might be grouped with folks who enjoy flicks about identity, man vs. man, those who aren't easily offended, etc. But there doesn't seem to be as clear a way to find a group of people who find aggression offensive, which is basically the driving theme of Fight Club. Perhaps given enough negative ratings it could be possible, but even though I've clicked 'Not Interested' on all the Disney movies, they keep suggesting I want their latest direct-to-DVD crapfest, so I'm left to assume they're rating mostly based on positive ratings.

Re:i was joking, however

[Pollardito]

Perhaps given enough negative ratings it could be possible, but even though I've clicked 'Not Interested' on all the Disney movies, they keep suggesting I want their latest direct-to-DVD crapfest, so I'm left to assume they're rating mostly based on positive ratings.

A co-worker gets almost no recommendations at all from Netflix, and customer service told him that they generate recommendations based on ratings of 4 or 5 (though you'd think that the recommendations that they do generate would have to filter through similar movies that you've rated at 0). He was told to rate the movies that he likes higher in order to fix it, but that's never really accomplished anything as he has several hundred movies in the 4-to-5 range and maybe a dozen recommendations total.

I'm pretty sure that the Disney/children's movie recommendation flood that most everyone seems to be getting is driven by parents who don't actually love those movies, but are rating those movies on behalf of their children. That causes a weird connection to movies that they themselves enjoy, and it makes it seem like the same audience is enjoying both types of movie. They need to have an "I'm a parent" flag somewhere to help them sort that out

Film recommendations

[michuk]

Does anyone find Netflix recommendations any good anyway? I used http://criticker.com/ for quite a while and was very happy about the recommended stuff. Recently switched to http://filmaster.com/ (which is a free service) and it's equally good, even though both probably use a pretty simple algorithm compared to Nextflix.

Re:Film recommendations

[coaxial]

I believe that Netflix is still using Cinematch. You could look into movielens. It's from the GroupLens group at U Minn.

[E]ven though both probably use a pretty simple algorithm compared to Nextflix.

You do know that Netflix said on the outset "You're competing with 15 years of really smart people banging away at the problem." and it was beat in less than week.

That's not to meant as a knock against Netflix's engineers, but more about that they didn't really build a state of the art recommender system. Simple SVD (aka latent semantic indexing outperfomed them as well.) They did something a bit more than straight up kNN clustering, but that was pretty much it.