DMSIG - Recommender Problems for Web Applications
LOCATION: LinkedIn, 2027 Stierlin Ct., Mountain View, CA 94043
DATE: Monday April 26, 2010; 6:30 - 8:30 pm
(6:30 - 7:00 networking & snacks; 7:00 - 7:10 announcements; 7:10+ presentation, Q&A)
COST: Free and open to all who wish to attend, but membership is only $20/year. Anyone may join our mailing list at no charge, and receive announcements of upcoming events.
SPEAKER: Deepak Agarwal is currently a Principal research scientist at Yahoo!
TITLE: Recommender Problems for Web Applications
DESCRIPTION: Several web applications like content optimization and online advertising involve recommending items from an inventory for each user visit to maximize some yield metric of interest (e.g. click rates). These are instances of large scale recommender system problems that entail several statistical challenges. We provide a mathematical description of the problem followed by modeling solutions for a content optimization problem that arises in the context of Yahoo! Front Page (www.yahoo.com). In fact, we discuss models to a) serve most popular items, b) serve items that are most popular in different user segments and c) provide personalized item recommendations for each user. Our models are based on time series methods, multi-armed bandit schemes and bilinear random effects model. One class of bilinear random effects model we propose extends reduced rank regression to incomplete matrices, the other class extends matrix factorization to incorporate covariates.
Throughout, concepts are illustrated with examples and results obtained from "bucket tests" conducted on a real system.