Title: Rank-consistency of the generalized Bradley-Terry model with link misspecification
Authors: Yongdai Kim - Seoul National University (Korea, South) [presenting]
Abstract: The estimation of ranks has received much attention in the areas of information retrieval and online game design, and the Bradley-Terry model is popularly used. We show that the popularity of Bradley-Terry model is due to not only its flexible modeling and easy computation but also its good asymptotic properties even when the model is misspecified. We provide the sufficient conditions under which the Bradley-Terry model yields a consistent estimate of ranks with partial preferences data when the true generative model of ranks belongs to the class of the Thurstone model. By numerical experiments, we illustrate that the proposed sufficient conditions are important and practically useful.