Three-Toed Sloth

"Personalized Content Recommendation on Yahoo!" (Next Week at the Statistics Seminar)

Attention conservation notice: Of limited interest if you (1) will not be in Pittsburgh on Monday, or (2) do not use the Web.

One of the first things I have the students in data mining read is "Amazon.com Recommendations Understand Area Woman Better Than Husband", from America's finest news source. The topic for next week's seminar is how to harness the power of statistical modeling to make recommendation engines even more thoughtful, attentive, delightful and broad-minded (all qualities for which statisticians are, of course, especially noted in our personal lives).

Deepak K. Agarwal, "Personalized Content Recommendation on Yahoo!"

Abstract: We consider the problem of recommending content to users visiting a portal like Yahoo!. Content for each user visit is selected from an inventory that changes over time; our goal is to display content for billions of visits to Yahoo! to maximize overall user engagement measured through metrics like click rates, time spent, and so on. This is a bandit problem since there is positive utility associated with displaying content that currently have high variance. Each user can be interpreted as a separate bandit but they all share a common set of arms given by the content inventory.

Classical bandit methods are ineffective due to curse of dimensionality (millions of users, thousands of content items to choose from). We take a model based approach to the problem and reduce dimension by sharing parameters across bandits and arms. In this talk, we describe latent factor models that capture interactions between users and content through multiplicative random effects model. We describe scalable methods to fit such hierarchical models through a Monte Carlo EM approach. Approximate model fitting in a Map-Reduce framework for massive datasets (that cannot fit in memory) is also described.

Time and place: 4--5 pm on Monday, 18 October 2010, in Doherty Hall A310

As always, the seminar is free and open to the public.

Enigmas of Chance