Pinterest's new retention model boosts engagement across Homefeed, Search, and more

Pinterest researchers have unveiled a new approach to optimizing recommender systems for long-term user engagement—without the usual complexity of custom reward engineering or reinforcement learning. The paper, accepted at Recsys 2026, introduces a model-agnostic downstream rewards learning framework that works across any ranking pipeline. The core insight: rather than directly modeling sparse, delayed retention signals, the team identified online session-level behaviors that occur soon after a recommendation and are highly predictive of a user returning. These behaviors become the reward signals, enabling standard machine learning models to optimize for retention without additional overhead.

The framework was productionized across Pinterest's major surfaces—Homefeed, Related Pins, Search, and Notifications—touching millions of daily active users. Online A/B experiments showed consistent increases in both short-term engagement metrics (clicks, saves) and longer-term retention rates. Crucially, the method is model-agnostic, meaning it can be dropped into existing recommendation models (e.g., DNNs, transformers) without re-architecting training loops. This practical, scalable technique could become a standard tool for any platform with recommendation systems that want to move beyond click-through rate optimization toward genuine user loyalty.