Towards Adaptive Mixed-Initiative Media Recommender Systems with LLMs

Recommender systems (RS) increasingly shape how people encounter digital media, but often leave users with limited understanding of why items are recommended and few ways to influence them. Existing user interfaces (UIs) typically offer static, one-size-fits-all explanations and controls that do not reflect users’ dynamic needs, limiting user agency. My research investigates how mixed-initiative recommender interfaces, enabled by large language models (LLMs), can adapt explanatory and interactive elements to users’ dynamic needs. I plan to develop a conceptual model that links situational media needs to requirements for information (understanding) and interaction (control). Building on this model, I will design and experimentally evaluate adaptive UIs that combine personalized explanations with interactive controls, measuring their effects on user understanding and perceived control. Through this doctoral consortium, I seek feedback on my research plan and guidance in designing effective experiments on adaptive mixed-initiative RS.