Adaptive Data Collection via Autoregressive Generation in News Recommendation Systems

Watch a 45-minute lecture from Columbia University professor Hongseok Namkoong at the Simons Institute exploring a novel framework for adaptive data collection algorithms using historical data. Learn how to leverage autoregressive models for predicting sequential feedback in real-world decision-making scenarios, with a focus on cold-start recommendation problems. Discover how the framework implements Thompson sampling with a learned prior by pretraining models to make accurate predictions based on action features and reward patterns. Understand the theoretical foundations showing how pretraining loss controls online decision-making performance, illustrated through a practical news recommendation system that incorporates fine-tuned language models processing article headlines. Explore solutions for handling uncertainty and active information gathering in changing environments where data availability is limited.