How Recommendation Systems Work At Netflix and Youtube: A Review of Candidate Generation, Ranking, Multitask Objectives and Responsible Personalization

Abstract

Streaming platforms such as Netflix and YouTube offer catalogues far larger than any user can browse, and most of what people watch is chosen from personalized recommendations. This paper explains how these industrial recommendation systems work, based on published research from the platforms and the wider community. It traces the evolution from the Netflix Prize and matrix factorization to deep learning, and describes the typical two-stage pipeline: candidate generation, which retrieves a few hundred relevant items from millions using embeddings learned from watch histories, and ranking, which scores candidates with richer features to predict clicks, watch time and satisfaction. It reviews multitask ranking that balances engagement and satisfaction objectives and corrects position bias, sampling-bias correction in two-tower retrieval models, reinforcement learning that optimizes long-term user value, and the personalization of rows, artwork and the home page at Netflix. Evaluation through offline metrics and online A/B testing is discussed, as are concerns about filter bubbles, calibration of recommendations to user interests and audits of problematic content pathways. A five-layer architecture is presented with research gaps, objectives, a comparison of techniques and challenges. The review concludes that modern recommenders are complex, multi-objective systems in which responsible design is as important as accuracy.