On YouTube’s recommendation system

What this post added

This post details the evolution and technical underpinnings of YouTube's recommendation system. It explains how the system has moved from a popularity-based "Trending" page to a personalized approach driven by user viewing habits. Key technical signals discussed include clicks, watchtime (introduced in 2012 to address clickbait), survey responses (to measure "valued watchtime"), sharing, likes, and dislikes. The post highlights the use of machine learning models trained on these signals to predict user satisfaction and the system's dynamic adaptation to changing viewing habits. It also touches upon the responsible recommendation aspect, including demoting sensationalistic content and predicting minors in risky situations, and the system's role in connecting users to high-quality information.