Deep dive into: Adapting the Facebook Reels RecSys AI Model Based on User Feedback

Delivering personalized video recommendations is a common challenge for user satisfaction and long-term engagement on large-scale social platforms. At Facebook Reels, we’ve been working to close this gap by focusing on “interest matching” – ensuring that the content people see truly aligns with their unique preferences. By combining large-scale user surveys with recent advances in machine learning, we are now able to better understand and model what people genuinely care about, which has led to significant improvements in both recommendation quality and overall user satisfaction.

Traditional recommendation systems often rely on engagement signals – such as likes, shares, and watch time – or heuristics to infer user interests. However, these signals can be noisy and may not fully capture the nuances of what people actually care about or want to see. Models trained only on these signals tend to recommend content that has high short-term user value measured by watch time and engagement but doesn’t capture true interests that are important for long-term utility of the product. To bridge this gap, we needed a more direct way to measure user perception of content relevance. Our research shows that effective interest matching goes beyond simple topic alignment; it also encompasses factors like audio, production style, mood, and motivation. By accurately capturing these dimensions, we can deliver recommendations that feel more relevant and personalized, encouraging people to return to the app more frequently.

To validate our approach, we launched large-scale, randomized surveys within the video feed, asking users, “How well does this video match your interests?” These surveys were deployed across Facebook Reels and other video surfaces, enabling us to collect thousands of in-context responses from users every day. The results revealed that previous interest heuristics only achieved a 48.3% precision in identifying true interests, highlighting the need for a more robust measurement framework. 

By weighting responses to correct for sampling and nonresponse bias, we built a comprehensive dataset that accurately reflects real user preferences – moving beyond implicit engagement signals to leverage direct, real-time user feedback.

Daily, a certain proportion of users viewing sessions on the platform are randomly chosen to display a single-question survey asking, “To what extent does this video match your interests?” on a 1-5 scale. The survey aims to gather real-time feedback from users about the content they have just viewed.

The main candidate ranking model used by the platform is a large multi-task, multi-label model. We trained a lightweight UTIS alignment model layer on the collected user survey responses using existing predictions of the main model as input features. The survey responses used to train our model were binarized for easy modelling and denoises variance in responses. In addition, new features were engineered to capture user behavior, content attributes, and interest signals with the object function to optimize predicting users’ interest-matching extent.

Analysis & Development

The UTIS model outputs the probability that a user is satisfied with a video, and is designed to be interpretable, allowing us to understand the factors contributing to users’ interest matching experience.

We have experimented with and deployed several use cases of the UTIS model in our ranking funnel, all of which showed successful tier 0 user retention metric improvements:

The UTIS model score is now one of the inputs to our ranking system. Videos predicted to be of high interest receive a modest boost, while those with low predicted interest are demoted. This approach has led to:

By integrating survey-based measurement with machine learning, we are creating a more engaging and personalized experience – delivering content on Facebook Reels that feels truly tailored to each user and encourages repeat visits. While survey-driven modeling has already improved our recommendations, there remain important opportunities for improvement, such as better serving users with sparse engagement histories, reducing bias in survey sampling and delivery, further personalizing recommendations for diverse user cohorts and improving the diversity of recommendations. To address these challenges and continue advancing relevance and quality, we are also exploring advanced modeling techniques, including large language models and more granular user representations.

.meta-btn { background-color: #0064E0; /* Meta Blue */ color: #ffffff !important; /* Force white text */ padding: 10px 20px; /* Button size */ border: none; /* No border */ border-radius: 5px; /* Rounded corners */ cursor: pointer; /* Pointer cursor on hover */ text-decoration: none; /* Remove underline */ display: inline-block; /* Button-like display */ font-weight: bold; /* Optional: bold text */ transition: color 0.3s ease, background-color 0.3s ease; /* Smooth transitions */ } .meta-btn:hover { color: #808080 !important; /* Grey text on hover */ }

Future Impact

Meta believes in building community through open source technology. Explore our latest projects in Artificial Intelligence, Data Infrastructure, Development Tools, Front End, Languages, Platforms, Security, Virtual Reality, and more.

Engineering at Meta is a technical news resource for engineers interested in how we solve large-scale technical challenges at Meta.

To help personalize content, tailor and measure ads and provide a safer experience, we use cookies. By clicking or navigating the site, you agree to allow our collection of information on and off Facebook through cookies. Learn more, including about available controls: Cookie Policy

Source: View Original