The ability of recommending good content items to users is key to the success of Web sites like Yahoo!. Usually, the set of candidate items are large and dynamic, and different users have different tastes. In this talk, I will give an overview of how latent factor models can be effectively used to predict the “rating” that a user would give to an unseen item, in order to recommend the items with high predicted ratings. In particular, I will begin with matrix factorization methods, which have shown excellent performance in the Netflix competition, where most users and items have sufficient past rating data. However, in many Web applications, recommendations have to be made when we have no or little past rating data for many items or users. I will address this “cold-start” problem by extending matrix factorization with feature-based regression, topic modeling and online learning.
Sunday, November 21, 2010
Free and open to the public