|
Abstract:
|
The problem of ranking from pairwise comparisons has been studied in multiple disciplines including statistics, operations research, computer science, machine learning, and others, and several algorithms have been proposed. These include, for example, least squares ranking, spectral ranking, maximum likelihood estimation under the Bradley-Terry-Luce (BTL) model, and others. However, not much has been understood about when these different algorithms perform well, and when they fail. In this talk, I will discuss recent work on developing a unified statistical perspective to understand the conditions under which different ranking algorithms succeed or fail. In particular, I will highlight the role of the underlying probabilistic model from which pairwise comparisons are assumed to be drawn, and will discuss how different algorithms succeed under different conditions on this probabilistic model.
|
