Symbolic Computation Group

Friday, November 12, 2010 at 10am, at U. of Western Ontario

Abstract:

How does one analyze a model generating millions of predictions every day? In this talk we will consider online ad auctions, paying particular attention to the ads optimized for conversions. These ads rely on probabilistic conversion prediction, which is carried out by extremely high dimensional models. We will take a simple metric used to analyze such models and show how it can be effectively computed from the vast click and conversion data with Sawzall, a programming language designed at Google for log analysis.