Conference Program

Recalibrating Probability Density Estimates Using Feature-Space Regression (310102)

Brett Andrews, University of Pittsburgh 
*Biprateep Dey, University of Pittsburgh 
Jeff Newman, University of Pittsburgh 

Keywords: Uncertainty Quantification, PDF Calibration

Uncertainty estimates from many machine learning methods do not fulfill the statistical definition of a probability density function (PDF).Previous works have used the global distribution of Probability Integral Transform (PIT) values to re-calibrate PDFs, but offsetting inaccuracies in different regions of feature space can limit the efficacy of those methods. We leverage a recently developed technique that characterizes the local PIT distribution at any location in feature space to perform a local re-calibration of PDFs. Though we focus on an example from astrophysics, photometric redshifts, our method can produce PDFs which are calibrated at all locations in feature space for any use case.