On YouTube’s recommendation system

This post provides a deep dive into the evolution and technical underpinnings of YouTube's recommendation system. It details the shift from a popularity-based 'Trending' page to a personalized system that leverages user viewing habits and compares them with similar users. Key technical signals discussed include clicks, watchtime (and its evolution to 'valued watchtime' using survey responses), sharing, likes, and dislikes. The post highlights the use of machine learning models to predict user satisfaction and the system's ongoing efforts to demote low-quality content and promote high-quality information, particularly for news and educational topics. It also touches upon user controls for privacy and the system's adaptation to different content categories.