Deep Neural Networks for YouTube Recommendations

The paper describes YouTube's recommendation system, one of the largest-scale and most sophisticated industrial recommenders in existence, at a high level while focusing on the dramatic performance improvements brought by deep learning. It follows the classic two-stage information retrieval dichotomy: a deep candidate generation model first narrows the enormous corpus, and a separate deep ranking model then orders the candidates.

Beyond the architecture, the authors provide practical lessons and insights derived from designing, iterating on, and maintaining a massive recommendation system with enormous user-facing impact. Sharing these production insights mattered because it connected academic deep learning methods to the realities of deploying recommendations at YouTube scale.