GenPage: Toward Netflix's End-to-End Generative Home Screen

Netflix's home screen goes beyond a simple list of recommendations, having a complex two-dimensional structure in which row composition, entity placement, and layout are all personalized. Existing approaches focused on listing individual recommended items, but this had the limitation of struggling to reflect the overall harmony and context of the entire screen.

GenPage introduces an Autoregressive approach to solve this problem. Instead of building the home screen all at once, this model uses the information from already-placed rows and entities as conditioning to sequentially generate the next row or element.

The core aspects of this technology are as follows:

• Context-awareness: The next element is determined by considering what is currently on the screen and what context of the user has been reflected.
• End-to-End generation: Rather than simply listing the outputs of individual recommendation models, the entire screen's structure is unified into a single generative process.
• Structural consistency: By learning the relationships between rows, a layout that is visually and contextually harmonious is constructed.