We introduce a transfer-learning method based on the Bradley--Terry model for multi-attribute pairwise-comparison data. The aim is to estimate the log-worth parameters of one primary attribute while using information from related secondary attributes. The method first pools the primary data with data from informative secondary attributes. It then corrects this pooled estimate using the primary likelihood, with a ridge penalty controlling the size of the correction. When the informative set is unknown, we use held-out primary data to select secondary attributes. For a known informative set and under a pooled Bradley--Terry compatibility condition, we derive high-probability $\ell_\infty$ and $\ell_2$ error bounds. Under additional conditions, these bounds can have a smaller asymptotic order than the corresponding primary-only Bradley--Terry bounds. We also establish asymptotic normality for a one-step estimator. A simulation study evaluates the method under more general settings, and an application to consumer preferences for eba, a cassava-derived food product, illustrates its use and interpretation. An R package implementing the method is available at https://CRAN.R-project.org/package=BTTL.

翻译：暂无翻译