CONCEPT Cited by 1 source
Whole-page optimization¶
Definition¶
Whole-page optimization is the practice of optimizing a recommendation or content layout as a single unit rather than scoring individual items independently. It captures cross-item interactions (diversity, stopping power, category balance) that entity-level objectives miss.
Why it matters¶
Traditional recommender stacks score entities independently (pointwise) or in pairs (pairwise). This ignores: - Diversity — a page of 10 thrillers may individually score high but collectively bore the user - Stopping power — a "Continue Watching" row placed early satisfies immediate intent but reduces browsing depth - Category balance — users want variety across genres, formats, content types
Implementation via autoregressive RL¶
Netflix's GenPage enables whole-page optimization by treating page generation as a sequential decision-making problem solved with reinforcement learning (Dr. GRPO algorithm): - Reward: Page-level sum of entity rewards (derived from user feedback) - KL penalty: Keeps policy close to pretrained checkpoint to avoid reward hacking - Emergent diversity: Homepage diversity increased during RL training despite not being in the objective — evidence that page-level optimization captures cross-entity interactions
(Source: sources/2026-06-29-netflix-genpage-generative-homepage-construction)