Sunday, March 23, 2008

Netflix Contest and Recommender Systems (short history)

I'm fascinated by the Netflix challenge: Netflix is offering one million dollars for an algorithm to improve recommendations based on movie ratings. Along with this go intermediate "progress" prizes of $50,000 per year the contest runs. The prize leaderboard is interesting reading, in that you can see who is entering teams, their results, and sometimes a bit about the team. Team Bellkor is a group of researchers from AT&T Labs, my first employer out of grad school. (I don't know these particular folks.) Netflix provides an interesting example of one company "funding" research at another company. The research lab folks are getting papers out of it, of course, probably the most important thing they need to do in a research lab (money comes second; although these days, proof that their work can impact a real business domain may beat everything else).

In another AT&T Labs connection, the Netflix prize FAQ cites an excellent overview paper on evaluation of recommender systems (pdf), co-authored by a colleague of mine, Loren Terveen, from the HCI department I was in at Bell Labs. Loren worked closely with Will Hill, one of the Bellcore researchers who (co-temporaneous with Pattie Maes at MIT and Paul Resnick) kicked off the work on recommender and ratings systems that you now find implemented all over the Internet. Recommender systems as a broad theme include all user ratings on products or comment postings (such as Amazon book ratings, or ratings implemented in almost all forum software now); they're intended to help others find good quality content by aggregates of ratings from other users, not from editorial oversight which is costly and therefore scales poorly to large amounts of content. There are important tweaks you can apply to your system or your filtering mechanism, such as "ratings of people like me" versus ratings of everyone, of course. (Netflix has some version of predicted "ratings for YOU", specifically, which I haven't investigated in any detail.)

I recommend glancing through Loren et al.'s paper, for a refreshingly meta perspective on a piece of technology that now defines a lot of assumptions behind what is called "web 2.0." As a more personal note, I wander among mostly non-research types these days, and the hot topics du jour (like "social networking") tend to get dropped into web system design discussions all the time, with a kind of naive "of course we need it" mentality. I can only sigh at how old I feel sometimes. Critical evaluation and careful implementation do matter, even for all the stuff that made it out of research projects into profit-making companies and community-platform toolkits.

As another personal note, I'm generally pleased by the level of researchy savvy I detect in the Netflix prize FAQ. Hey, if you're hiring at a software company, consider investing in some serious research-minded folks for competitive advantage!

1 comment :

Anonymous said...

Thanks for the links! I've been watching the Netflix contest story from afar and want to learn more about it.