Every student will complete a project. There are two options for the project:
Research projects can be either theoretical or experimental. We have less guidance for experimental projects, but if you have concrete proposals, you can run them by the instructor.
For a theoretical project, a good way to start is with an in-depth read of a paper. As you're reading the paper, think about what you might want to do differently. Question the authors' assumptions. Some of the assumptions might be too strong. Try to see if you could still show some interesting theoretical results under weaker (or somewhat different, but interesting) assumptions. Also, maybe the *kinds* of guarantees the authors showed are different from some type of guarantee you would like to show? Also, are there potentially missed connections in the paper (there seems to be a related idea, but the authors either didn't notice this idea or didn't mention it). What happens if you pursue one of those connections? Also, theoretical projects do not have to exclude experimental approaches. You might run experiments (perhaps simulations) to help guide you in your theoretical pursuit.
For an experimental project, you also might start with an in-depth study of an existing empirical paper and see what new questions arise. Then, pursue some of those questions. Consider how the authors were able to run experiments. Did they use simulations? Can you also do simulations? Alternatively, did they use real data? Within the timeframe of the project, would you be able to obtain suitable data for your experiments? It is important to remember that experiments should be guided by theory. What hypothesis do you have? What questions are you experimentally trying to answer. Having concrete, testable questions in mind helps give structure (and justification) to experimental pursuits.
The other option for the project is an in-depth study of a specific topic. For that topic (for example, peer prediction), you should do a review of the literature on that topic, looking at many (or at least several very good) papers on the topic. Your goal should be to figure out the state of progress. What are the best (or most interesting) results achieved so far? Are there negative results (things that are known not to be achievable)? Have people in the area identified open problems? Which ones are still open? Finally, the end goal of your in-depth study should be to identify research questions that you could tackle. In that sense, you can think of this in-depth study of the topic as setting yourself up to find a great research question to work on.
For any pairs of students, either both students must be graduate students, or both students must be undergraduate students.
Project proposals will be due around the mid-point of the semester. The proposal should be about 1 page. You should describe the topic/problem you chose, explain why it is important, and tell us what you hope to achieve. Also, indicate a few of your next steps. We want to see that you have clarity about what you are going to do.
Project presentations will be in last week of class (or possibly in the last two weeks of classes). All students are required to attend all project presentations.
Your final report will be due some time after the last class, in the early part of the final exam period.
💡 A lot of the work in machine learning with strategic agents assumes that agents best-respond to an algorithm’s current hypothesis by manipulating their features. What happens if agents can strategically decide whether to participate or not? How should one frame a good learning objective? How can one design a good learning algorithm here?
đź’ˇ In performative prediction, people often view the goal as minimizing the performative risk. Is this really the right objective? The paper Alternative Microfoundations for Strategic Classification suggests that the performative risk can be problematic. How could you think of a better objective and, moreover, how to design a good algorithm for this objective?
💡 Bayesian games form a widely studied and important class of games. There are several extensions of the concept of correlated equilibrium to Bayesian games, and people in the field do not agree about which extension is the “right” one. A good starting point could be the paper “Bayes correlated equilibria, no-regret dynamics in Bayesian games, and the price of anarchy” (Kaito Fujii, COLT 2025). What are the different definitions of Bayes correlated equilibria? For each one, is it a good solution concept? Why or why not?
đź’ˇ Crowdsourcing can be modeled as principle-agent problem, where a principle (a company operating a crowdsourcing platform) uses a mechanism to pay crowdworkers. There are a variety of desiderata of interest here, such an wanting agents to both exert effort and to be truthful (not labeling or predicting based on how an agent thinks others might label or predict). What are some of the fundamental techniques that have been devised here, and what are their strengths and weaknesses?
đź’ˇ In crowdsourcing, how can a mechanism be designed so that it works well in the face of human crowdworkers that might be surreptitiously using an LLM? How could such use be disincentivized without penalizing honest workers? There may be recent papers in this area but also a lot of room for introducing new ideas.
đź’ˇ In political polls, a forecast of an event is made, and agents (like political parties) may use the forecast to affect the outcome being forecast. What are good ways to think about information elicitation mechanisms in this type of setting? How could you formally set up the problem? A good starting point could be the paper "Incentivizing honest performative predictions with proper scoring rules" by Oesterheld et al., but there are many different perspectives one could take on this problem.