A1210
Title: Modeling pairwise comparison data with cyclic and acyclic structures under a Bayesian framework
Authors: Hisaya Okahara - Tokyo University of Science (Japan) [presenting]
Tomoyuki Nakagawa - Meisei University (Japan)
Shonosuke Sugasawa - Keio University (Japan)
Abstract: Pairwise comparison data frequently arises in diverse applications such as sports analytics, preference learning, and human feedback evaluation. Classical models, such as the Bradley-Terry model, assume that preferences can be explained by latent scores associated with each item, leading to a globally consistent ranking. However, real-world data often exhibit cyclic or intransitive patterns that cannot be captured by score-based approaches alone. A statistical modeling framework that extends classical pairwise comparison models is introduced to flexibly accommodate such complex patterns while preserving interpretability. The approach is formulated under a Bayesian logistic regression setting, where a Gibbs sampling framework enables efficient posterior computation. Finally, through numerical experiments, the proposed framework is illustrated to be capable of capturing cyclic structures observed in practice.
