🔵Meta Engineering·January 14, 2026

Enhancing Recommendation Systems with User Feedback at Scale

This article details Meta's approach to improving Facebook Reels' recommendation system by integrating direct user feedback through a User True Interest Survey (UTIS) model. It highlights how moving beyond traditional engagement signals to understand true user interests leads to better personalization, engagement, and content diversity in large-scale ranking systems. The architectural integration of the UTIS model into different stages of the ranking funnel is a key system design aspect.

AI & ML Infrastructure Distributed Systems Performance & Scaling

Read original on Meta Engineering

Meta's Facebook Reels team has evolved its recommendation system by addressing the limitations of relying solely on implicit engagement signals like likes and watch time. These traditional metrics often capture short-term value but fail to reflect users' true, nuanced interests. To overcome this, they developed the User True Interest Survey (UTIS) model, a direct feedback mechanism designed to capture deeper user preferences.

Challenges with Traditional Recommendation Systems

Traditional recommendation systems, while effective at scale, often optimize for readily available engagement signals. This can lead to: 1) Over-optimization for viral, but not necessarily relevant, content. 2) Difficulty in surfacing niche or high-quality content that might not immediately generate high engagement. 3) A lack of understanding of the 'why' behind user interactions, making it harder to truly personalize experiences.

The UTIS Model: Architecture and Integration

The UTIS model is built on large-scale, randomized in-app surveys asking users about the interest-match of a recently viewed video. This sparse, but high-quality, feedback is used to train a lightweight UTIS alignment model. This model takes existing predictions from the main multi-task, multi-label ranking model as input features and outputs a probability of user satisfaction. The integration of this 'perception layer' into the existing large-scale ranking system is crucial for its effectiveness.

Integration Points in the Ranking Funnel

Late Stage Ranking (LSR): UTIS acts as an additional input feature into the final value formula, allowing for fine-tuning based on true interests while balancing other ranking concerns.
Early Stage Ranking (Retrieval): UTIS helps reconstruct and re-rank user interest profiles, enabling the system to source a more relevant set of candidates from the vast content pool. Knowledge distillation is used to align large sequence models with UTIS predictions from LSR.

💡

System Design Takeaway: Hybrid Recommendation Approaches

This case study demonstrates the power of combining implicit signals with explicit user feedback. In designing large-scale recommendation systems, consider how to architect for hybrid models that can adapt to evolving user preferences and overcome the biases inherent in purely behavioral data. This often involves integrating lightweight, specialized models into various stages of a complex ranking pipeline.

recommendation systemsmachine learninguser feedbackranking systemspersonalizationlarge-scale systemsA/B testingsystem architecture

Comments

Loading comments...

Architecture Design

Design this yourself

Design a large-scale video recommendation system that incorporates both implicit user engagement signals and explicit user feedback through a survey mechanism, considering how to integrate the feedback model into different stages of the ranking and retrieval pipeline.