Extraction of Latent Variables for Modeling Subjective Quality in Time-series Human-Robot Interaction

This study presents a novel method for modeling subjective evaluation of the quality of interaction (QoI) by extracting explanatory variables that are not explicitly quantifiable by humans from human-robot behavior. The proposed method extracts latent variables that account for both explicit and tacit knowledge by performing maximum likelihood estimation to predict manually selected explanatory variables, alongside QoI score prediction, from time-series interaction data. In this study, we address three key questions: (i) whether the extraction of latent variables improves accuracy compared to conventional regression analysis, (ii) whether implicit variables, beyond those selected by humans, play a significant role, and (iii) whether human-selected explanatory variables are necessary in explaining subjective assessment scores. The results of comparisons across several learning conditions demonstrate that incorporating tacit knowledge variables, uncorrelated with traditional explanatory variables, enhances the accuracy of QoI estimation. This study contributes by enabling data-driven extraction of explanatory variables, revealing the influence of tacit knowledge on QoI estimation, and highlighting the importance of both top-down and bottom-up approaches in accurately estimating subjective evaluations of QoI.