Modern recommendation systems, such as Netflix or Spotify, are designed to predict what items to recommend to someone based on information such as their previous watch history, what the people most similar to them have liked, or other forms of implicit or explicit feedback. While this has demonstrably been effective for providing watchable recommendations, it also encourages poor content consumption habits and leaves users with few options for self-reflecting and little ability to alter their recommendations in a manner that is clear and direct. In this talk, I present our work exploring the potential affordances of representing our content preferences as natural language profiles and examining how we evaluate such recommendation systems.

Biography of the speaker: Joyce Zhou is a doctoral candidate in the Computer Science Department at Cornell University. Before Cornell, they got a B.S. and M.S. in Computer Science from the University of Washington. They have contributed to research on human-AI interaction, collaboration, explainable AI, and text visualization at scale. Their current research interests include human-interpretable and steerable recommendation systems, taking advantage of modern developments in language models. Outside of research, Joyce has taught AI as a visiting course instructor at UPenn, held workshops on introductory CS concepts for grade schoolers, and entertains their cat in their free time.