Tuesday, October 03, 2006

Netflix Asks $1 Million Question

Netflix is offering $1 million to whomever can figure out a better way to recommend movies to its customers. The company wants to improve its recommendation (aka collaborative filtering) system for suggesting movies based on movies that other people liked, according to CNet. The company is making a huge database of movies ratings available to contestants.

But answering this challenge requires overcoming almost insurmountable obstacles -- that people have unique tastes, and that movies are not that easy to categorize. No two people agree on the majority of things in life (ask any married couple) such as their taste in clothes, cars, or whom they consider attractive. For example, just because two people like Bruce Willis and Jim Carrey movies doesn't mean they will also agree on Charlie Chaplin.

Also, films can be broken into countless sub-genres that makes it difficult to associate taste from one person to the next. A simple star rating system doesn't suffice -- the more important question is what about the movie did the person like, breaking it down to its component parts.

The same difficult question of associating taste has been asked by fashion designers, record companies and realtors, but the answer requires getting to know each individual better rather trying to do statistical analysis across a large data set.

Posted By John Gartner at 10:46 AM
Permanent Link: Netflix Asks $1 Million Question | Comments (1)

(1) Comments on Netflix Asks $1 Million Question

The interesting thing about collaborative filtering is that across a large enough data set, it does capture these things that are so hard to quantify in the typical relational database/quantified parameters fashion.

It is often tough to capture enough data. For example, "people who rented X also rented Y" is not nearly as useful as "people who loved X also loved Y" which, in turn, is not nearly as useful as "people who loved X, *and* hated Y, *and* who think Jim Carrey's only good movie is the Truman Show, also loved C...or also hated B."

Movie critics have always served as a very crude kind of collaborative filtering for me. For example, if the guys at Ain't It Cool really hate a film, then chances are I will, too. If they really like a film, then odds are good I will also, unless the film is a comic book film or it is directed by Guillermo Del Toro, in which case I have to discount their enthusiasm and simply chalk it up to them being fanboys. Doesn't mean the comic book or GDT movie will stink, but I can't use the Ain't It Cool recommendation as a reliable data point in the positive direction.

So, it seems to me that the key to improving on a system like Netflix's collaborative filtering is to get users to give you more and more information...possibly by giving a discount on the service fees for filling out extended surveys, or providing extra demographic information, etc. Of course, it could also be really creepy for Netflix to know that muh about me.

I am already paranoid of what my Tivo must think of me...I already know it thinks I have a thing for Mark Harmon and Tom Selleck. Is it my fault I enjoyed both Summer School and Quigley Down Under?!?!?

Comments by Clegg : Tuesday, October 03, 2006 at 08:59 PM

Post a Comment

Name:
Email:
WebSite:
Comment:
Remember Me
Subscribe