New Framework Boosts Long-Term User Engagement in Recommender Systems

Recommender systems traditionally focused on short-term user interactions, but the industry is shifting towards optimizing for long-term engagement and retention. Directly measuring retention is challenging due to sparse and delayed feedback signals. Existing solutions often involve complex, task-specific reward engineering and significant computational overhead, making them difficult to generalize across different platforms. Researchers have developed a new, unified framework designed to optimize long-term user value in large-scale recommendation systems. This model-agnostic approach learns downstream rewards by first identifying session-level behaviors that are both observable early in a user's journey and highly predictive of their future retention. It then derives various reward signals from observed user action patterns across diverse sources. The framework has been successfully deployed across multiple Pinterest surfaces, including Homefeed, Related Pins, Search, and Notifications. Online A/B experiments have consistently shown improvements in key engagement and retention metrics, demonstrating its practical effectiveness and scalability.