Dr. Matteo Sesia
Assistant Professor of Data Science and Operations,
USC, Marshall School of Business
Matteo Sesia received a B.S. in Engineering Physics and a M.S. in Physics of Complex Systems from Politecnico di Torino (Torino, Italy), a M.A. in Statistics and Applied Mathematics from Collegio Carlo Alberto (Torino, Italy), and a Ph.D. in Statistics from Stanford University. After completing his Ph.D. in 2020, Dr. Sesia joined the faculty of the Marshall Business School at the University of Southern California as an Assistant Professor of Data Sciences and Operations.
WATCH LIVE: November 2nd at 12:30 pm
Abstract: Sophisticated machine-learning algorithms are increasingly relied upon for a variety of applications, despite the relative lack of guarantees on their out-of-sample performance. This talk will focus on the particular problem of rigorously quantifying the uncertainty of predictions computed by multi-class classification algorithms, leveraging simple ideas from statistical exchangeability that yield a practical method with useful finite-sample guarantees. In particular, by building upon the existing work on conformal prediction and other data hold-out methods, we describe novel conformity scores that can be used to produce uncertainty estimates with desirable theoretical oracle properties as well as good empirical performance. Experiments on synthetic and real data will demonstrate the practical value of our theoretical guarantees, as well as the statistical advantages of the proposed methods over the existing alternatives.