A modified Gaussian process regression is proposed aiming at making regressors robust against outliers. The proposed method is based on U-loss, which is introduced as a natural extension of Kullback-Leibler divergence. The robustness is examined based on the influence function, and numerical experiments are conducted for contaminated data sets and it is shown that the practical performance agrees with the theoretical analysis.
ASJC Scopus subject areas
- Physics and Astronomy (miscellaneous)