Abstract: We study the course allocation problem, where the aim is to optimally allocate bundles of courses to university students. Although the number of course bundles is exponential in the number of courses, prior mechanisms rely on the unrealistic assumption that they have access to the full ordinal preferences of the students. Recent experimental work has confirmed that this assumption harms the efficiency of these mechanisms. To address this limitation, we propose a new machine learning-powered course allocation mechanism. The main goal of integrating machine learning into a course allocation mechanism is to improve preference elicitation, which is the main challenge, given the exponential bundle space and the students' mistakes when reporting their preferences. Concretely, we extend the well-established Course Match mechanism with an iterative machine learning-based elicitation procedure. We show that our design maintains the attractive theoretical properties of Course Match in terms of welfare and fairness. Finally, our experimental results suggest that our approach outperforms Course Match in terms of efficiency and fairness in several realistic settings.
