Uncertainty Quantification in Learning from Pairwise Comparisons

Speaker: Ye Zhang(University of Pennsylvania)

Date: May 27 Fri. 1:30pm

Venue: Tencent Meeting Room: 758 543 870

Abstract: Ranking from pairwise comparisons is a central problem in a wide range of learning and social contexts, and the Bradley-Terry-Luce (BIL) model is one of the most studied models for ana1yzing ranking data. Despite all the recent progress, uncertainty quantification under the BIL model remains unclear when only a small number of comparisons is observed. To address this challenge, we first establish non-asymptotic entrywise distributions of the maximum 1ikelihood estimation and the spectral method under the BIL model. We then develop statistical inference, procedures for individual rankings and preference parameters.