I was introduced to machine learning by Microsoft Research researchers who were working in the Bing search engine. I think I was one of the first, if not the first, leading economist to say, "Hey, this is really big, and I think that there are things about machine learning that you can bring into economics that would really make things different."
- Susan Athey, Stanford Economics of Technology Professor
Susan Athey was one of the first “tech economists.” She served as consulting chief economist for Microsoft Corporation for six years, and she is currently a Consulting Researcher with Microsoft Research New England. Professor Athey recently sat down with Eric Horvitz, director of Microsoft Research, to share her enthusiasm for the value economics brings to the internet and artificial intelligence.
Below are a few excerpts from Fireside Chat with Susan Athey. The entire conversation is available as a video from Microsoft Research.
Combining Theoretical and Empirical Work
My interest in theoretical issues has always been as a means to an end. Early in my career, I was interested in studying problems about auctions and mechanism design, and then there were some technical issues that needed to be solved to really make that theory beautiful. In some ways it was a little bit unnatural for me because I got more theoretical. And I became very well known for very theoretical work and I loved doing it and it was beautiful; but actually, it wasn't my main motivation. My main motivation was to solve these slightly more applied problems as a means to an end. The class of problems I was interested in were games under uncertainty, auctions and mechanisms. And really designing markets: trying to figure out how you can design institutions or rules of games to make them work better and work more efficiently.
One example from timber auctions: We looked at the difference between having an open ascending auction where people raise their hands until the auction ends and everybody drops out, versus the first price sealed-bid auction where you send in sealed bids and then the winner is the highest bidder.
Under some conditions, there's a beautiful theorem that says that those are actually equivalent. But under real world conditions, there are several reasons they're not equivalent. In particular, the ones where you have sealed bids are inefficient, that is a weaker bidder can win. It's basically that these people shade their bids; and weaker bidders are more aggressive bidders – they can outbid someone who thinks they're likely to win and doesn't think they need to be so aggressive. That seems like that's bad and you shouldn't do that; but it has this extra benefit in that it induces more participation of weak bidders. So the bidder, understanding that they have a chance to sneak in and win, are more likely to participate. And that in turn can fight collusion and lead to more competitive auctions.
I did a lot of research about that theoretically and also working with data from real timber auctions to illustrate the magnitudes of those effects. Those types of themes, the rules of the game, have direct consequences but also the broader implications for how competitive a market is and who even participates in the market. This is a theme that I brought later to the advertising auctions at the search engine at Microsoft.
Machine Learning and Economic Modeling
This is still fairly niche in economics, but I feel like it's on the verge of exploding. I was introduced to machine learning by Microsoft Research researchers who were working in the Bing search engine. I think I was one of the first, if not the first, leading economist to say, "Hey, this is really big, and I think that there are things about machine learning that you can bring into economics that would really make things different."
Economics has always been good at counterfactual prediction, cause and effect, what-if questions. But always in a small data framework. Roughly you have three covariates and a million observations. We worked on flexible models, but in a very small number of covariate setting. … Data mining was a pejorative term, that was like what people did that were cheating.
What I feel we've really imported very well from machine learning into economics is this data driven search for models, and much richer classes of models, and much better, more systematic ways to fit the data well.
That was really why I got excited about working on AI – why I thought I had something to bring to the table – because all of my training was about counterfactuals, but it was applied to very simple data sets and really simple problems. With these huge new data sets, as well as opportunities in tech firms to intervene and run lots of experiments, there was a possibility to learn about cause and effect in a much grander scale. And so, I want the theoretical insights from people like Judea Pearl or Don Rubin, that set up a framework for thinking about this. Those theoretical insights can be implemented in a much richer way when you combine the computational techniques and the estimation techniques from machine learning to do things like personalization and to control for lots of covariates, but still maintaining the rigorous theoretical focus on cause and effect.
As an example, in old-style economics we might have said, "Let's look at the average effect of changing a price." And then we might run an experiment with price changes, or it might be that there's some natural experiment in the past, something that caused this price to change that was unrelated to consumer demand. But we would estimate an average of that. Machine learning has gotten really good at personalized predictions. So, you can adapt a lot of the algorithms from personalized prediction or just personalized fitting, like saying, "Okay, which of these people is going to click?" or "Which of these people is going to buy?" and then adapt those techniques to answer the question, "For which of these people will a price change cause them not to buy or cause them to buy?"
Taking a Career Risk
I was a professor at Harvard. I had just gotten this big prize. [Professor Athey received the John Bates Clark Medal in 2007.] Everybody loved what I had done to date. At that point, I had defined the field as market design, which was bringing theory and data together to solve auction problems, and school choice problems, and things like that. I was well recognized as the leader of my generation.
But then the opportunity came to come to Microsoft. I just was just seduced by this idea of... I went in and I looked at the search engine, and it was just wide open. I realized right away that my economic tools were not up to the task, actually. This was harder, and broader, and more complicated. But at the same time, I had the instinct that there really was economics here and there was value to add.
The above is a small sampling of the topics discussed in Microsoft Research’s Fireside Chat with Susan Athey. Enjoy the entire conversation between Professor Susan Athey and Eric Horvitz.
Susan Athey is the Economics of Technology Professor at Stanford Graduate School of Business. She also is Professor of Economics (by courtesy), School of Humanities and Sciences, and Senior Fellow, Stanford Institute for Economic Policy Research. Her current research focuses on the economics of digitization, marketplace design, and the intersection of econometrics and machine learning. She has worked on several application areas, including timber auctions, internet search, online advertising, the news media, and virtual currency.
Professor Athey was awarded the Jean-Jacques Laffont Prize in Toulouse, France in 2016. The annual award recognizes an internationally renowned economist whose research is in the spirit of the work undertaken by Professor Jean-Jacques Laffont, combining both the theoretical and the empirical. In 2007, Professor Athey was named the first female recipient of the American Economic Association’s prestigious John Bates Clark Medal, awarded every other year to the most accomplished American economist under the age of 40 “adjudged to have made the most significant contribution to economic thought and knowledge.”
Note: The Technology | Academics | Policy (TAP) website is sponsored by Microsoft Corporation. While Microsoft provides administrative and financial support for the site’s platform and content, there is no payment made to scholars for appearing or blogging on the site.