Netflix's GenPage: Generative AI for Homepage Construction

Shifting from Multi-Stage to Generative AI in Recommendations

Traditionally, systems like Netflix's homepage recommendations rely on complex, multi-stage pipelines involving separate components for candidate generation and ranking at both row and entity levels. GenPage represents a significant architectural departure, adopting a single generative transformer model inspired by Large Language Models (LLMs). This model treats user history and request context as a prompt and autoregressively generates the entire homepage as a response, including rows, entities, and layout simultaneously. This contrasts with most generative recommenders that produce flat ranked lists, allowing for holistic page construction rather than isolated scoring.

Architectural Motivations and Benefits

• End-to-end modeling: Consolidates multiple ML models into a single transformer, reducing maintenance overhead, eliminating misaligned objectives across stages, and minimizing feature engineering.
• Whole-page optimization via Reinforcement Learning (RL): Enables optimization for page-level rewards, capturing complex interactions (e.g., diversity, stopping power) across rows and entities that are difficult to model with entity-level objectives.
• Better scaling behavior: Provides a clearer path to quality improvements by leveraging more data, compute, and model capacity inherent in transformer architectures.
• Flexibility and extensibility: The prompt-response paradigm simplifies support for new content types (games, live events), diverse layouts, personalized UI components, and artwork personalization with fewer architectural changes.