Brian pointed out that the AT&T Research Labs team that won this year’s Netflix Progress Prize ($50k out of $1M) for improving movie recommendations had published a number of papers on their winning strategy. It’s interesting reading, and this paper is fairly approachable if you skip the statistics in the middle.
Their final approach combined 107 different model, and though the majority provided only incremental improvement, the total effect propelled them to an 8.43% improvement over Netflix’s own proprietary algorithms. (Wikinomics fan can take a moment to cheer another success for open collaboration.)
One interesting sidenote:
The distribution of movies-per-user is quite skewed. Figure 4 shows this distribution (on a log scale). Ten percent of users rated 16 or fewer movies and one quarter rated 36 or fewer. The median is 93. But there are some very busy customers, two of which rated over 17,000 of the 17,700 movies!
This confirms my previous findings on participation and the 1-9-90 rule. The team also made use of additional models which considered simply the presence or absence of a rating for a movie from a particular user.
Overall, it seems that Netflix and the recommendation community have gotten a lot of mileage out of the prize in an area that will continue to grow:
Because good personalized recommendations can add another dimension to the user experience, e-commerce leaders like Amazon.com and Netflix have made recommender systems a salient part of their web sites.
If such systems interest you, I’d recommend O’Reilly’s Programming Collective Intelligence. Reading - and working through it’s examples - really opens your eyes to how simple these algorithms can be and how commonplace they’ve become.
