Netflix initial recommendation system

Netflix introduced a single foundation model to replace many separate recommender systems, allowing one unified architecture to power multiple personalization tasks. The model tokenises rich user interaction data—such as watch events, duration, device signals, and metadata—so it can learn long-term behavioural patterns. With techniques like sparse attention and sliding-window sampling, it efficiently handles long histories at scale. It also solves cold-start issues by combining learned embeddings with metadata-based representations, enabling recommendations for new titles instantly. Overall, the approach improves recommendation quality, scalability, and system maintainability across the platform.