How Netflix Uses Matrix Factorization and Tiered ML to Power Recommendations

Netflix's recommendation engine relies on a technique called Matrix Factorization, which breaks down a massive user-movie ratings grid into smaller matrices to predict how much any user will enjoy a title they have never watched. Because recalculating this for over 250 million users in real time is impractical, Netflix splits its machine learning pipeline into three tiers: offline batch processing on Spark or Hadoop clusters, nearline asynchronous updates triggered by user events, and an online real-time layer that serves ranked results within milliseconds. The system also uses Contextual Bandits to dynamically personalise thumbnail artwork based on a viewer's watch history, aiming to boost click-through rates. For example, the same film may display a romantic scene to one user and a comedic moment to another, depending on their inferred preferences. Ultimately, Netflix's goal across all these systems is not just prediction accuracy but long-term user retention.