Leveraging LLMs for Explainable Recommender Systems: Exploring User Perceptions and Faithfulness in Generated Explanations
Abstract
Recommender systems (RSs) are widely used to support users in discovering relevant items, yet their growing complexity raises concerns about transparency and interpretability. This work investigates the use of Large Language Models (LLMs) to generate automated, human-centered explanations for RSs and evaluates whether such explanations are faithful to the models’ underlying decision processes. First, we conducted a user study assessing personalized movie recommendations and explanations generated by GPT-3.5 Turbo, measuring perceived effectiveness, personalization, and persuasiveness. Then, we evaluated the faithfulness of explanations across movie, music, and book domains using four LLMs. Results show that LLM-based recommendations increase user satisfaction compared to random baselines; however, explanations frequently fail to align with the factors driving the recommendations. Moreover, user-based explanations were not consistently perceived as superior to generic ones.