Deep Neural Networks for YouTube Recommendations

This paper describes YouTube's recommendation system, one of the largest-scale and most sophisticated industrial recommender systems in existence, at a high level and focuses on the dramatic performance improvements brought by deep learning. It is organized around the classic two-stage information retrieval dichotomy: a deep candidate generation model first selects candidates, and a separate deep ranking model then orders them.

Beyond the model architecture, the authors provide practical lessons and insights derived from designing, iterating on, and maintaining a massive recommendation system with enormous user-facing impact. The paper is notable for detailing how deep learning was applied at industrial scale within a real production recommender serving an enormous audience.