ResearchPod Summary
Modern recommender systems often struggle to balance short-term engagement (e.g., clicks) with long-term user retention. While short-term signals are easy to measure, they are often noisy and can lead to feedback loops that prioritize shallow, repetitive browsing. Conversely, direct optimization for long-term retention is difficult because the feedback is sparse, delayed, and hard to attribute to specific recommendations. This paper addresses this gap by identifying and leveraging "downstream rewards"—early, observable session-level behaviors that are statistically predictive of future user return.
To identify which behaviors actually drive long-term value, the authors developed an offline screening framework. They analyzed user activity data to distinguish between "shallow" browsing (which often correlates with lower retention once exposure is controlled) and "deep" engagement. The framework uses a Random Forest model to evaluate potential reward candidates against three criteria: correlation with future revisitation, observability within a session, and incremental predictive power. This analysis revealed that deep P2P (Pin-to-Pin) exploration, saves, and diverse cross-category engagement are the most robust indicators of long-term retention.
The authors propose a unified, model-agnostic approach to incorporate these insights into existing ranking systems. They define three primary reward families:
These rewards are designed to be flexible, allowing them to be integrated into various deep learning-based recommendation models across different surfaces like Homefeed, Search, and Notifications. By providing denser, less delayed feedback than traditional retention labels, these signals help models adapt more effectively to evolving user interests.
This framework provides a practical, scalable solution for large-scale platforms to shift from optimizing for immediate clicks to optimizing for sustainable user value. By moving the optimization objective closer to the user's actual long-term goals, the authors demonstrate that it is possible to improve retention without the massive computational overhead typically associated with complex reinforcement learning implementations.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.