Proper losses for means, with connections to prediction markets
Rafael Frongillo (Berkeley)
NICTA SML SEMINARDATE: 2012-03-22
TIME: 11:15:00 - 12:00:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
Proper losses are loss functions whose minimizers are "well-calibrated" in a probabilistic sense. While typically defined for multiclass classification problems, I will consider the general problem of learning summary statistics (in a broad sense) of distributions. I will present a new characterization of proper losses for means, and leverage this characterization to draw a strong connection to prediction markets.
