2026 Volume 76 Issue 4 Pages 146-163
Pairwise comparison-based evaluation methods are effective when it is difficult to perform a direct evaluation of multiple items. Traditionally, businesses have conducted analyses on a small number of products. However, with the expansion of e-commerce, comparing and selecting from among dozens of options has become increasingly common. The number of required pairwise comparisons increases quadratically with the number of items. When evaluations are conducted manually, practical applications are generally limited to around 1,000 comparisons. Thus, conventional methods that rely on human raters to provide pairwise comparison data for all possible combinations become infeasible. Some methods have been proposed to reduce the required number of comparisons by leveraging the structure of the pairwise comparison matrix to impute missing data. However, these approaches typically achieve only a modest reduction, and their applicability becomes limited when a large proportion of data is missing. As a result, they remain insufficient as a solution to the increasing number of evaluation targets. To address this issue, this study proposes a subjective evaluation method based on pairwise comparisons that remain applicable even when the number of evaluation targets is large. This approach utilizes machine learning to impute missing entries in the pairwise comparison matrix. Specifically, the present study focuses on the relationship between item features and subjective evaluation scores, and training a machine learning model to estimate missing pairwise comparisons based on auxiliary information. Through experiments using synthetic and real-world datasets, the present paper demonstrates the proposed evaluation model's effectiveness and practical applicability for analysis.
