Abstract
Learning with inexact supervision, rather than definite labels, has been proposed to relieve the labeling burden. Pairwise comparison (Pcomp) is a novel inexact supervision setting for binary classification where only pairwise samples with positive confidence priority are provided. Despite the existence of unbiased risk estimators (UREs) for the Pcomp classification, the absence of definite labels inevitably yields a decline in accuracy. In the current study, we propose a multi-view Pcomp classifier, aiming to mitigate the accuracy decline using easily accessible multi-view feature representations when definite labels are not available. Specifically, a hinge-style upper bound of the original URE is theoretically obtained, minimizing which as a loss function, makes it feasible to implement the Pcomp classification. Subsequently, a privilege multi-view Pcomp classification model (PMV-Pcomp) is proposed, in which view-agreement and cross-view privilege constraints are integrated to promote consensus and complementary principles. A theoretical generalization error bound for PMV-Pcomp is also presented to guarantee its reliability. Moreover, noise sample pairs with incorrect positive confidence priorities may provide inappropriate guidance to the model. Therefore, a robust version of PMV-Pcomp (RPMV-Pcomp) is designed that filters out suspected noise sample pairs using a dual-model denoising mechanism before training. Exciting numerical results of comparative experiments and ablation study validate the competitive generalization ability of our RPMV-Pcomp.
View more >>