Hello from Vegas: StyleFeeder hits KDD 2008
I’m representing StyleFeeder at this year’s KDD conference, held in Las Vegas, Nevada. It might seem odd mixing seductive showgirls and stodgy statisticians, but I think it’s an excellent location choice. Gambling concepts such as probability, expected value and exploration vs. exploitation are core to many concepts in Machine Learning, Data Mining and Statistics.
KDD played host to the 2nd Recommender System/Netflix Prize Workshop. Gavin Potter showed us that users, movies, and even ratings sessions (date) impart significant biases on ratings, so much so that a model which simply captures these biases and completely ignores user-movie affinity yields a lower error score than than the original CineMatch algorithm. After some discussion of the fact that minimum-error recommenders tend to yield popular and somewhat uninteresting recommendations, Oscar Celma and Pedro Cano presented a study of this effect on music. They found that a collaborative filtering similarity metric was strongly biased toward popular music, whereas content-based and expert-based similarity metrics made it easier to explore “the long tail.” Next, a member of the Gravity Team, Gabor Takacs (who I later learned is the author of the best “big board” tic-tac-toe player in the world), provided a detailed description of their methods for the Netflix Prize. Their approach is an SVD-like matrix factorization, which incorporates incremental training, regularization, user/item bias, positivity constraints, and neighbor-based correction.
Based on discussions and other presentations, it sounded like a combination of matrix factorization and neighborhood based methods was the most common successful approach to the Netflix Prize competition. Everyone at the workshop seemed to agree that Netflix did a surprisingly good job of selecting a goal for the competition: Netflix requires a 10% improvement over their CineMatch algorithm and the current top team has a 9.15% improvement. The difference seems small enough that the 10% goal must be reachable, but progress has slowed considerably, with improvement of only .72%-age points since the first progress prize was awarded last October.
As the main conference has started, it has become quite clear what the “hot” topic of the year is: social network modeling. Sessions on the topic have been packed and some top figures in the community have presented papers on the subject…