Machine learning algorithms now play a major role in all kinds of decision-making scenarios, such as college admissions, credit approval, and resume screening.
When the stakes are high, self-interested agents --- about whom decisions are being made --- are increasingly tempted to manipulate the machine learning algorithm, in order to better fulfill their own goals, which are generally different from the decision maker's.
This highlights the importance of making machine learning algorithms robust against manipulation.
The main focus of my research is on designing and analyzing machine learning algorithms that are robust against strategic manipulation, which is different from the relatively well-studied notion of adversarial robustness.
My research sets the foundations for several key problems in machine learning in the presence of strategic behavior:
Empirical risk minimization and generalization in classification problems:
Traditional wisdom suggests that a classifier trained on historical observations (i.e., an empirical risk minimizer) usually also works well on future data points to be classified.
Is this still true in the presence of strategic manipulation?
Distinguishing distributions with samples:
Due to various constraints, often we have to judge the quality of a data point based on a few samples (e.g., screening job candidates based on a few representative papers).
How should we calibrate our judgment when these samples are strategically selected or transformed?
Planning in Markov decision processes:
Dynamic decision-making problems (traditionally modeled using Markov decision processes) can be solved efficiently when the decision maker always has complete and reliable information about the state of the world, as well as full control over which actions to take.
What happens when the state of the world is reported by a strategic agent, or when a self-interested agent may interfere with the actions taken?