Academics / Courses / DescriptionsCOMP_SCI 497: Prediction for Decision-Making
VIEW ALL COURSE TIMES AND SESSIONS
Prerequisites
CS PhD students or permission of instructorDescription
In this seminar course, we will consider what it means to make good predictions for the purpose of assisting decision-making. Powerful prediction models are increasingly available in domains like medicine, finance, and criminal justice. However, there is often a mismatch between the ways in which these models are developed–e.g., to minimize population level loss–and the needs of decision-makers who could benefit from the models. For example, how can we best deploy AI/ML models when the quality of individual decisions are important in a domain, such as when we must be able to ensure that a specific decision was not biased or obviously inferior to some alternative decision based on the same information? What do we gain by accounting for knowledge about the downstream decision task in training a model, and how do we best incorporate this information in the modeling-to-decision pipeline? When and how can we predict whether additional information or abilities a human decision-maker may have over model predictions will lead them to outperform a learned decision rule?
This course will cover recent and classic attempts to characterize good prediction for decision-making. Topics will include human versus algorithmic judgment, model-assisted human decision making, uncertainty quantification and calibration, individualized treatment rules, and algorithmic fairness.
- This course satisfies the Project or Technical Elective.
Coursework: Students will prepare and lead discussions on the papers selected. Coursework includes a midterm essay and a preliminary study for a research project.
INSTRUCTOR: Prof. Jessica Hullman