The 2016 RecSys paper "Deep Neural Networks for YouTube Recommendations" introduced a two-stage architecture for large-scale recommendation systems, moving from matrix factorization to deep learning. It utilizes a candidate generation model to narrow down options and a ranking model to predict expected watch time, optimizing for user engagement using implicit feedback. Read the full paper at research.google.com .