Slashdot Mirror
Build a Better Netflix, Win a Million Dollars?
An anonymous reader writes "In a quest to better movie recommendations, Netflix is opening their database (nytimes, registration and first child required) to users to try to craft a better recommendation technology. The problem is not easy. Says one researcher: 'You're competing with 15 years of really smart people banging away at the problem.'" Recommender systems are really an interesting problem, and that is likely very interesting data to play with.
if(user.getGender()==Person.MALE)
recomendation=MovieGenre.PORN;
else
recomendation=MovieGenre.CHICKFLICK;
And of course, slashdot must have sensed my post as my image word is "pervert"
..except, instead of making it open to the community (which is not a bad idea, I must say) I thought of having Google do it. This is, perhaps, IMHO, a much better idea. Now, what we really need is a Movie Genome Project, much like the Music Genome project that lead to Pandora.
Zagreus sits inside your head, Zagreus lives among the dead, Zagreus sees you in your bed and eats you in your sleep.
As a NetFlix user I have one suggestion for their recommendation system that can make it much better. Make it aware of the connection between series. That is to say, If you rent season 1 of something, suggest season 2, not season 4 (even if season 4 has better review ratings). If I mark season 1 of something as "not interested" instead of giving it a user rating, don't suggest every other season of that same show at the top of my recommendations. I mean how many times do I have to tell you I don't want to see any season of "Friends" ever, even if you pay me?
Why is it that the Slashdot editors are just too damn lazy to look up the RSS feed links to these pages?
While this may be true, I wouldn't let it deter you. Collaborative filtering is a field that is far from dead. The interesting thing about collaborative filtering is that on the surface, it seems pretty straight forward but once you dig into the mechanics of it, there is actually a lot of playing you can do. Ironically, the way you display the data to the end user is often what determines how well of a job you did.
Allow me to take a naïve approach at this topic and say we generate a movie index of each person. I would have A Clockwork Orange and Koyaanisqatsi at 5 while The Ring 2 would be at the very low end. My friend might have similar movies. If he has A Clockwork Orange up there, you might be able to compute a Euclidean distance between us. However, this approach falls apart because no one has seen Koyaanisqatsi and of the 20 movies I've ranked highly, they are hard to find.
You don't have to stop there, however. You could also database the movies I marked as "uninterested" or the movies that were presented to me but I didn't vote on. Like if I had seen the offer to mark J-Lo's latest flop but didn't, wouldn't that tell you something about me?
So these caveats present themselves all along the way and, at the end computation, you have many different strategies for this data. For example, while you might not be able to link my friend an I through movies, how far apart are we on a nod network? What I mean is, if you plotted every user in their own dimension depending on the movies they ranked and attempted to compute as good a distance as possible between all users, how far would I be away from my friend by hopping on these nodes? There's a lot of information to be gleaned in this sort of friend-of-a-friend collaborative approach.
Now you need to present this information to the user. Do you just up and recommend him a movie? Do you take Amazon's approach and say "Other people did this -- so should you."? Or do you give them some sort of three dimensional flash plotting of you versus the people nearest to you? Do you allow the user to contact those closest to them? Those farthest away?
My point is that while 15 years of research has been done, it doesn't mean there's been 15 years of testing and implementation which, in the end of creating products, is where most of the importance lies.
My work here is dung.
The problem with recommendation systems is that they use too little information to catagorize their subject.
What they need to do is copy the methods of the Music Genome Project (www.pandora.com), and list a larger set of attributes for the films. This way it can recommend films by checking many more characteristics, such as director, tone, writer, or subject.
http://www.netflixprize.com/
I saw the Sign, and it opened up my eyes
I wish they'd fix the problems in the logic determining what they actually send me from my queue before fixing problems with what they recommend to me. If I've got season 1 of a show in my queue prior to season 2, don't start sending me season 2 because some disc of season 1 is unavailable (which has happened to me multiple with both netflix and blockbuster online), send me something else completely. They've got the tech to keep one season of a tv show in order, it can't possibly be that difficult to extend that to keeping multiple seasons of a show in order.
On top of that, don't show me that it's available in my queue but send me something else instead. While I haven't asked netflix about this, I have asked blockbuster online, and I imagine they are both doing the same thing. The disc is "available" just not at the warehouse used to ship to me personally. Instead of basing one piece of information off of total stock and one off of local stock, base them both on the stock at the warehouse shipping to me.
Any marketer will tell you that what people tell you they want and what people actually want are very different things. Even if people answer honestly, the data you gather is often unreliable: people simply don't have as good a handle on what they want as they think they do.
Not that marketers have a better handle, but simply that people will swear up and down that they would buy a peanut-butter-filled hot dog, that they loved the one they tried, and then don't actually buy any.
Don't believe me? Go see Snakes on a Plane. Nobody else did. (Sure, $33 million seems like a lot, but that's chump change for a major studio release these days.)
The best improvements will come from insights gained between the lines. You may have rated The English Patient eleventeen stars, but if your next seven rentals were all episodes of The Girls Next Door, which you only rated 3 stars, it certainly looks like you want more Hugh Hefner and less Ralph Fiennes.
The best data is the data that the subject doesn't realize he's giving you. Once you start imposing conscious choice on the ratings, you get only what they say they like, not what they really like.