- We’ve enhanced personalized video recommendations on Facebook Reels by incorporating direct user feedback instead of solely relying on metrics like likes and watch time.
- Our new User True Interest Survey (UTIS) model helps showcase niche, high-quality content, thereby increasing engagement, retention, and satisfaction.
- We are advancing personalization by addressing issues like sparse user data and bias, and leveraging advanced AI for smarter, more diverse recommendations.
- Our paper, “Improve the Personalization of Large-Scale Ranking Systems by Integrating User Survey Feedback” provides full details on this work.
Providing personalized video recommendations is crucial for user satisfaction and long-term engagement on large-scale social platforms. At Facebook Reels, we aim to ensure content aligns with users’ unique preferences through “interest matching.” By merging large-scale user surveys with the latest in machine learning, we now better understand and model user preferences, significantly boosting recommendation quality and user satisfaction.
Why True Interest Matters
Traditional recommendation systems typically use engagement signals – likes, shares, watch time – but these can be noisy, not accurately reflecting users’ true interests. Models relying solely on these signals might maximize short-term user value but miss the mark on sustaining long-term product utility. To address this, we needed a more nuanced approach to measure content relevance. Effective interest matching extends beyond topic alignment to include audio, production style, mood, and motivation. Capturing these nuances allows us to offer users more relevant and personalized recommendations, enhancing app engagement.
How We Measured User Perception
To validate our approach, we initiated large-scale, randomized surveys within the video feed, asking users, “How well does this video match your interests?” These surveys, conducted across Facebook Reels and other video surfaces, helped us gather numerous in-context user responses daily. Previous interest heuristics had only reached 48.3% precision in identifying true interests, underscoring the need for a stronger measurement system.
We’ve built a comprehensive dataset, adjusting responses to correct for sampling and nonresponse bias, moving beyond implicit engagement signals toward direct, real-time user feedback.
<img decoding="async" class="aligncenter wp-image-23429" src="https://allyoucantech.com/wp-content/uploads/2026/02/adapting-the-facebook-reels-recommendation-system-using-user-feedback-1.png" alt width="277" height="600" srcset="https://engineering.fb.com/wp-content/uploads/2025/11