Colloquium Explores the Connections Between Computer Science and Economics

Internet users are all too familiar with the bright advertisements that blink in the web browser’s margins. They ask us to check our credit scores, suggest we apply for mortgage loans, and instruct us to buy flowers for Mother’s Day. But few of us realize that behind each ad is a complicated system of technology and economics to ensure the timing, pricing, and placement is perfect.

R. Preston McAfee, director of Google Strategic Technologies, discussed the economics of online advertising during McCormick’s first CS+X colloquium. Students, faculty members, and staff packed the Ford ITW Auditorium on April 3 eager to hear McAfee’s talk “Machine Learning in an Exchange Environment.”

“I have a theory,” McAfee said, “that the advertising systems are the largest statistical estimations that have ever been done.”

McAfee said that online advertising is also one of the most complicated practices. Advertisers only get paid when the ad is clicked—not just when it is seen. “As a result,” McAfee continued, “in order to say what’s the best ad to run on a given page, I need to forecast the probability that there’s a click.”

Much of what goes into the forecast relies on targeting. Advertisers track the types of people who visit certain sites, analyzing their demographics, geographical location, marital status, interests, and more. They use these variables when competing with other advertisers in auctions for ad space hosted by website publishers. The value of the ad is the winning bid multiplied by the forecasted click-through rate.

Machine learning in economics involves very large regressions with as many as a billion explanatory variables. Due to the scale, point estimates are typically used, and common economic concepts are ignored. The overwhelming amount of data makes it difficult to create and run accurate prediction models.

“We have this giant regression and all these different variables,” McAfee said. “The question is: ‘What should we do right now?’”

He says advertisers should adjust their estimates on the value of the ad to take into account the fact that they might learn something. By placing a bid in the auction, understandings of the parameters will get sharper. “Data from today is incorporated into estimates for tomorrow,” McAfee says. “This gives us a different perspective.”

Hosted by the Department of Electrical Engineering and Computer Science, the CS+X series explores ways in which computer science can augment and be augmented by other disciplines.

“The interaction between computer science and economics has a great future,” McAfee said. “The fruit is low on the tree. It’s a great discipline to be in.”

McCormick will host Luca Cardelli, principal researcher at Microsoft Research, for the next CS+X colloquium on Tuesday, April 29.