Yahoo! Research has a lab in New York city, right near Times Square. It is where I spent last year as as a "computing innovation fellow," and as promised, I want to blog about my experience.
Yahoo!'s New York lab isn't large. It has about 15 scientists, whose research spans several areas of computer science — machine learning, algorithms, economics and computation, storage systems, and computational social science. Most of the scientists are permanent researchers, but there also are a few postdocs. The lab often had talks, visitors, and interns to keep the atmosphere lively and interesting.
I mainly worked with John Langford on machine learning problems in computational advertising. A particular problem most search engines need to solve is what advertisements to show to their users, given some contextual information like a user's IP address, shared browser settings, and search query. The goal is to show users ads that they are likely to click, in order for the user to have a good experience and for the search engine to make a profit. However, how to do that is not clear because whenever an ad is shown, no explicit information is received about how well a different ad would have performed in the same situation. Effective algorithms need to balance exploiting strategies they have already learned to be good and exploring new strategies, possibly at some cost. In my time at Yahoo! we made good progress in understanding what sort of guarantees are possible for this problem, which is called the "contextual bandit problem" in the machine learning literature.
It was quite exciting to get to work on real-world internet scale machine learning problems, and have the results of my work have practical consequences. I was also very lucky to have a chance to work with John, who, in addition to being one of the foremost experts on this problem (and in machine learning in general), was a kind and patient host, from whom I learned a lot. I also got to work with Rob Schapire (my undergraduate mentor and collaborator thereafter), who was visiting the lab on his sabbatical from Princeton.
Overall, I was impressed by the flexibility scientists at Yahoo! have in choosing their problems, in having significant time to do research, and in having access to large amounts of real-world data at scales available at only a few other places. We computer scientists are quite lucky to have industrial research jobs as a serious option for research careers.
Now, I have started on a new adventure as a postdoc at Georgia Tech's Algorithms and Randomness Center. I have never been at such a big and "happening" department, and I am excited to get to work with many great researchers on interesting and important problems.