Progressive Comparison for Ranking Estimation

Abstract

Object ranking is a problem that involves ordering given objects by aggregating pairwise comparison data collected from one or more evaluators; however, the cost of object evaluations is high in some applications. In this paper, we propose an efficient data collection method called progressive comparison, whose objective is to collect many pairwise comparison data while reducing the number of evaluations. We also propose active learning methods to determine which object should be evaluated next in the progressive comparison; we propose two measures of expected model changes, one considering the changes in the evaluation score distributions and the other considering the changes in the winning probabilities. The results of experiments using a synthetic dataset and two real datasets demonstrate that the progressive comparison method achieves high estimation accuracy with a smaller number of evaluations than the standard pairwise comparison method, and that the active learning methods further reduce the number of evaluations as compared with a random sampling method.