Explanations are crucial for improving users' transparency, persuasiveness, engagement, and trust in Recommender Systems (RSs) by connecting interacted items to recommended items based on shared attributes. However, evaluating the effectiveness of explanation algorithms regarding those goals offline remains challenging due to their subjectiveness. This paper investigates the impact of user-level explanation properties, such as diversity and popularity of attributes, on the user perception of explanation goals. In an offline setting, we used metrics adapted from ranking to evaluate the characteristics of explanations generated by three state-of-the-art post-hoc explanation algorithms, based on the items and properties used to form the explanation sentence, across six recommendation systems. We compared the offline metrics results with those of an online user study. The findings highlight a trade-off between the goals of transparency and trust, which are related to popular properties, and the goals of engagement and persuasiveness, which are associated with the diversification of properties displayed to users. Furthermore, the study contributes to developing more robust evaluation methods for explanation algorithms in RSs.

翻译：暂无翻译