21 April 2008

Netflix Recommendations

Netflix is offering a million dollars to anyone who can supply a recommendation system that's 10% better than their currently deployed CineMatch algorithm. Example of why they are paying big $$$ for such an improvement:


They need all the help they can get.

It's an interesting problem to tackle. How do you identify the traits of a movie? Genre is the first that comes to mind, but certainly more is required to recommend a movie. Simple classification such as "Comedy" doesn't necessarily mean I will like Superbad if I enjoyed Tenacious D: The Pick of Destiny.

Based on the Netflix website, you can see each movie profile page has several specific categories, or sub-genres. I imagine the current CineMatch system uses these as a tag recommendation algorithm. For example, the page for There Will Be Blood lists these genres:

Drama
Indie Dramas
20th Century Period Pieces
Dramas Based on Classic Literature
Period Pieces
Dramas Based on the Book

As you can see, each sub-genre has its own landing page. If two given movies have a similar sub-set of genre tags and I liked one of them, it stands that I might enjoy the other. Let's go back to the image example up top...






Orange CountyBSG: Season 1Walk Hard
ComedyTelevisionComedy
Teen ComediesTV Sci-Fi and FantasySpoofs and Satire
--TV DramasContemporary Movie Musicals
----Blu-Ray


Other than the "Comedy" link, do YOU see any other similarities? I don't.

OK, what else can movies be related on? What about cast and crew? Ah-HA! Orange County and Walk Hard have the same director and both feature actor Jack Black. Are those truly related criteria, though? Jack Kasdan also directed indie crime thriller Zero Effect. If Jack Black was in Zero Effect, would CineMatch have recommended it, too?

Again, movie recommendation is an interesting problem. I wonder if it could ever work like music recommendation engine Pandora. Check out what I mean by reading this page at How Stuff Works. Could we identify such specific things as plot elements as seen on IMDB.com pages? If I liked a movie with a memorable "peacefully gazing at the night sky" scene, could be there be any recommendation significance to a movie with a similar scene?

Anyone want to go for the million bucks with me? =)

p.s. If anyone can find one single similarity between BSG Season 1 and Walk Hard/OC, please comment!

6 comments:

  1. pps - there is ZERO whitespace or code inbetween the paragraph and the table of genre tags in the blogger code of this post. where the heck is all that whitespace coming from?

    ReplyDelete
  2. if you win a million dollars, will you take care of the bills for a while?

    ReplyDelete
  3. Maybe Dewey Cox is a Cylon?

    I've had some unusual "If you liked this..." results with Amazon.com too. Somehow PSPs and Family Guy are related.

    ReplyDelete
  4. i think that a big component of their current system is looking at numbers of people that rated those highly. So my guess is there happens to be a more than statistically significant portion of the netflix population who gave 4 stars or more to those three movies, and who also match many of your other favorite movies. I think it works off of similarity in rating as much as it does off of movie descriptions, where as pandora works all off what you like and not what other people liked too.

    ReplyDelete
  5. ah, you're suggesting that they relate users to each other and recommend based off of similarly rated movies. rather than relate movies to each other and recommend from there. this puts a greater recommendation accuracy on users who supply more feedback about their likes and dislikes. users who put more into the system get more out of it, too! interesting...

    ReplyDelete
  6. I think that there are certain movies that, if given ANY kind of good rating, should result in the renter losing their Netflix membership! For example, Battlefield Earth, Judge Dredd, Evolution, etc. What other films are so bad that I can use them to screen out people with no taste?!

    ReplyDelete