We’re now well into the last third of my Networks course, which focuses on the interaction of economics and networks. We talk about a lot of different topics, but one that the students really seem to enjoy is ad auctions, which tends to take up the last three or four lectures of the class — and also serves as the context for our last competition / mini-project: clickmaniac.
I find teaching about ad auctions really fun because, for many of the students, it’s the first time they’ve really thought in depth about the challenges of designing good pricing mechanisms and auctions…and it’s fun to let them explore those topics with various thought exercises during class. Though, it’s funny how mysterious the whole space of computational advertising is for them. I always ask early on how many students block ads, and this year all but 2 folks in class said that they do! So, they’re learning about an enormous industry that they avoid almost entirely…
Another fun thing about teaching this topic at Caltech is that I can really point to Caltech as being fundamental to the emergence of the big ideas in the space. The first company to use auctions to price search ads was Overture (now part of Yahoo), which involved a lot of Caltech folks and spun out IdeaLabs, which is an incubator next door to Caltech that is very tied in with the faculty here. (At least two of our CS faculty are doing startups there currently.) And, in addition to the Caltech connections with Overture, Caltech alum Gil Elbaz was a leader in the development of AdSense, which underlies Google’s system. So, it’s certainly a space in which Caltech folks have made a huge impact.
Theory vs. practice in ad auctions
I find that teaching about ad auctions is kind of a challenge because of the stark contrast between the simple theoretical stories we tell about auctions (VCG and GSP), versus the reality of participating in these auctions. In particular, as the CS theory view of ad auctions basically derives from a “standard” econ view of multi-item auctions, which highlights that VCG is the natural choice for a design and that, in some sense, the GSP auctions that emerged at places like Google were “flawed” because of their lack of incentive compatibility.
But, of course, when one considers the real systems, there are so many other factors, e.g. budgets and learning clickthrough rates, that you have to really squint pretty hard to see the theory. It becomes even harder when you start to consider the broader landscape of ad exchanges, information providers, etc.
So, when I teach this topic, I really try to have some exercises that highlight the contrast between the simple theoretical models we use to reason about ad auctions and the reality. For example, this Friday, the whole class will be devoted to a live “simulation” of ad exchanges where students take on the roles of information providers and advertisers, and I play the role of the auctioneer. Once they finish playing the game, which was designed by Eric Bax at Yahoo Research, they really start to have a feel for the complexities of the interactions between information gathering, learning, and bidding…and they really start to see the huge value of information for targeted bidding — something that is completely hidden by the classical models that we use to prove theorems about ad auctions.
In addition to this game, I also provide them the opportunity to actually play as advertisers in reality in the clickmaniac competition…
Crowning a Clickmaniac
Thanks to Facebook, which sponsors clickmaniac, we give groups of students $3 a day to spend on advertisments, and ask them to try to target & bid in order to maximize the number of clicks and likes they get for our course facebook page. The ads they are using are all targeted for charities (which were chosen with input from the students in the class), and our facebook page then has links to the charities, so that by the end of the competition we’ve had a positive impact beyond simply learning about ad auctions…
Of course, despite the fact that Facebook uses VCG under the hood, figuring out how to bid and target in these auctions is quite complex, so the students have a lot of fun figuring out which strategies work well and which groups are effective to target… I’ll be very curious to see how effective they are this year (and how much money we raise for the charities).
As with rankmaniac and pandemaniac, while there is a competition to see which team is the clickmaniac, the grade for the assignment comes from getting more clicks than a few benchmarks (which are set by the TAs).
If you’re curious about the details, you can see the writeup here.