From latent factors to language: a user study on LLM-generated explanations for an inherently interpretable matrix-based recommender system

User perception matters more than automatic metrics: The study of 326 participants found that traditional automatic evaluation metrics often fail to capture what users actually need and value in explanations, highlighting the importance of directly asking users about their experiences with AI systems rather than relying solely on computational measures.

LLMs can effectively translate mathematical models into understandable language: The research demonstrates that large language models can successfully convert the internal workings of an interpretable recommendation model (based on matrix factorization) into natural language explanations that users find transparent and satisfactory, bridging the gap between technical interpretability and user comprehension.

Different explanation strategies produce only modest differences in user perception: While the study tested multiple explanation types by varying the information provided to the LLM, all explanation approaches were "generally well received" with only "moderate statistical differences between strategies," suggesting that users may be relatively forgiving of variation in how explanations are presented.

Multiple dimensions of explanation quality matter to users: The research evaluated explanations across five distinct dimensions—transparency, effectiveness, persuasion, trust, and satisfaction—indicating that responsible AI systems should consider how explanations perform across these varied user concerns, not just whether they increase trust or acceptance.

Qualitative user feedback provides crucial insights beyond numbers: User comments in the study offered "complementary insights beyond the quantitative results," demonstrating that responsible AI evaluation should combine numerical ratings with open-ended feedback to fully understand how people experience and react to AI explanations.