Seminar, Hosted by Fields Institute for Research in Mathematical Sciences
Transfer conformal predictive inference for regression
University of Alberta
Abstract
With a small target dataset, conformal prediction for regression can yield intervals too wide to be informative. Two transfer-learning algorithms borrow data from related sources, including when the useful sources are unknown or differ in distribution from the target. Conditional Kullback–Leibler divergence guides source selection. Finite-sample bounds, simulations and real datasets show when intervals can narrow while preserving coverage.