UC Santa Barbara

This study explores the integration of Embodied Conversational Agents (ECAs) with Conversational Recommender Systems (CRS), focusing on the impacts of visual representation and recommendation bias. Leveraging an open-source Large Language Model alongside Nvidia Audio2Face and Unreal Engine 5, this study developed ECAs to assess their influence on user interaction in CRS. A 2x2 between-subjects study with 53 participants examined interactive avatars versus animated icons, evaluating their roles in enhancing recommendation persuasiveness and shaping user decision-making. Results indicate that while interactive avatars significantly boost user engagement, their influence on recommendation persuasiveness is minimal. Conversely, the presence of recommendation bias within CRS significantly impacts user opinions, highlighting its crucial role in CRS design. These findings underscore the importance of balancing visual innovation with ethical recommendation strategies in CRS, offering insights for advancing user-centered system development and informing future research.