How Netflix Built a Single Model for Search and Recommendations

Discover how Netflix revolutionized their machine learning infrastructure by consolidating fragmented ML pipelines into a single, unified architecture serving over 300 million users in this 47-minute conference talk. Learn about Netflix's transition from bespoke ML systems to UniCoRn (Unified Contextual Ranker) and their proprietary User Foundation Model, which leverages Transformer architectures similar to GPT-4 to understand and predict user behavior trajectories. Explore the technical challenges of scaling personalization systems, including the two-stage ranking framework that balances latency, throughput, and service level agreements while maintaining high-quality recommendations. Examine how Netflix's team approached tokenizing user history by treating titles like words in natural language processing, enabling their foundation model to capture complex user preferences and viewing patterns. Understand the system design considerations for building ML infrastructure at Netflix's scale, including strategies for addressing over-personalization and filter bubbles that can limit content discovery. Gain insights into the measurable impact of unified personalization systems and learn practical approaches to reducing technical debt while increasing personalization effectiveness. The presentation includes detailed discussion of fine-tuning strategies, cold start problems for new users, and integration of multi-modal data sources, making it essential viewing for engineering leaders, ML architects, and anyone interested in large-scale recommendation system design.