World's Fair 2025
Netflix's Big Bet: One model to rule recommendations: Yesu Feng, Netflix
Overview
Netflix is betting on a single foundation model, based on transformer architecture, to handle all recommendation use cases. This approach aims to improve personalization by scaling semi-supervised learning and create high leverage by integrating the model across all systems, simultaneously enhancing downstream applications. The model learns user representations and incorporates rich event data, addressing challenges like cold-start problems and enabling faster innovation.
Who should watch
- AI Engineers
- Product Managers
- Builders working on recommendation systems
- Those interested in large-scale AI model deployment
- Engineers focused on infrastructure consolidation and innovation velocity
Key takeaways
- Netflix developed a foundation model to unify diverse recommendation needs across different content types (movies, TV shows, games, live streaming) and page layouts (homepage, search, mobile feed).
- The previous approach involved many specialized, independently developed models, leading to duplicated feature and label engineering, making it unscalable and hindering innovation.
- The foundation model leverages transformer architecture and incorporates learnings from LLMs, including careful tokenization of user interaction events, which have multiple facets beyond simple IDs.
- Key components of the model include event representation (time, location, device, canvas), embedding layers combining ID and semantic content information, transformer layers for user representation, and a rich objective layer supporting multitask learning.
- Scaling laws apply to recommendation systems, with Netflix observing continuous gains by increasing model parameters and data size, though latency requirements necessitate distillation.
- Learnings borrowed from LLMs include multi-token prediction for robustness and long-term satisfaction, multi-layer representation for stable user profiles, and long context window handling for efficient training.
- The foundation model is integrated into downstream systems via subgraph integration, by pushing out content and user embeddings, or by fine-tuning for specific applications.
- The big bet has been validated, showing high leverage in both AB test wins and infrastructure consolidation, enabling faster innovation and easier scaling to new applications.
Notable quotes
"The question we asked at that time was can we centralize the learning of user representation in one place."
"The answer is yes and we had this key hypothesis that about foundation model based on transformer architecture."
"We see we indeed see high leverage of FM to bring about both AB test wings as well as infrastructure consolidation."
Unofficial community note. Prefer the recording for nuance.