- Facebook Reels’ friend bubbles highlight Reels your friends have liked and interacted with, making it easier to discover new content and connect over common interests.
- This article describes the technical architecture behind friend bubbles, including how machine learning estimates the strength of relationships and ranks the content your friends interact with, increasing opportunities for meaningful engagement and connection.
Friend Bubbles enhance your social experience on Facebook Reels by helping you discover content your friends enjoy, create shared viewing experiences, and spark new conversations. A quick tap on a bubble will start a one-on-one conversation with a friend using that reel.
This feature combines social and interest signals to recommend more relevant, personalized content and make it easier to start conversations with the people who matter most. When videos connect to both personal and friend-related interests, it creates a feedback loop that improves recommendations and strengthens social connections.
Friend Bubbles System Architecture Overview
The Friend Bubble Recommendation System includes several components that work together to display relevant content that your friends have interacted with by blending video quality signals with social graph signals.
- Intimacy between viewer and friend (interaction is most important): Identify which friend interactions are most likely to interest your audience.
- Video relevance (which videos to show): Rank videos that are contextually relevant to your audience.
Multiple friends interacting on the same video often indicates stronger common interests and higher relevance. Content that surfaces through connections with friends also tends to be of high quality, creating a reinforcing loop where social discovery increases engagement, which in turn further strengthens the social graph.
Viewer-Friend Intimacy: Identifying Friends Using a User-User Intimacy Model
Friend Bubbles relies on two complementary machine learning models to identify which connections people feel closest to. One model is based on user research feedback. The other is based on interactions on the platform.
Survey-based intimacy models leverage a wide range of signals, including social graph features (mutual friends, strength of connections, interaction patterns) and user attributes (behavioral and demographic signals such as user-provided location, number of friends, and number of posts shared) to build a more complete picture of real-world relationships.
It is trained on a regular cadence using a lightweight binary survey that asks a randomly selected group of Facebook users whether they feel close to a particular connection in real life. The survey is structured as a predictor of intimacy vs. non-intimacy, updated regularly to keep the labels current, and includes questions that serve as a proxy for offline relationship strength (e.g., how often two people communicate). In production, the model runs inferences on trillions of person-to-person connections between Facebook friends every week.
While survey-based intimacy provides a strong foundation, Friend Bubbles also uses a context-specific intimacy prediction model trained on activity signals on the platform, using the actual interactions that occur when a bubble is displayed (e.g., likes, comments, re-shares). This allows the model to capture in-context intimacy, that is, how likely a viewer is to value content recommended by someone in their friend graph based on their interactions on the platform.
Our approach emphasizes connection quality over quantity. Bubble adoption naturally increases as your friend graph grows, but more bubble video views don’t necessarily increase user engagement. The goal is to surface relevant friend connections (friend connections that are most likely to make the social context meaningful) by combining existing familiarity signals with surface-specific features that better reflect the relationship dynamics behind friend-driven recommendations.
Video Relevance: Make your ranking system friend content conscious
We use two key strategies to ensure that high-quality friend-interacted content reaches our users through the recommendation funnel. It’s all about expanding the top of the funnel and allowing the model to effectively rank friend bubble content through a continuous feedback loop.
Inventory Sourcing: Expanding the Top of the Funnel
In the search stage, we source candidate videos based on close friends identified by the intimacy model above. By explicitly capturing content that your friends have interacted with, you extend the top of the funnel and ensure sufficient candidate volume for downstream ranking stages. This is important because without it, high-quality friend content may not make it into the ranking pipeline in the first place.
Enable your model to effectively rank your friends’ content through a continuous feedback loop
A key insight from our development process was understanding why videos of interacting with friends sometimes struggle to rank well. That’s not because the video quality was poor, but because the model lacked the context of intimacy between users. Without that context, the model cannot learn what makes friends’ content uniquely valuable: that its relevance is often driven by relationship strength and social meaning, rather than the same signals that explain interest in content more generally.
To address this gap, we integrated friend-bubble interaction signals as a feature and added a new task to a multi-task, multi-label (MTML) model for both early- and late-stage rankings to incorporate viewer-friend relationship strength and learn downstream engagement in videos with social bubbles. By adding these signals throughout the ranking funnel, models can better recognize the value of content interacted with with friends, learn the relationship between intimacy and audience interest, and rank the most relevant, high-quality friend content higher.
The system includes a continuous feedback loop where friend and bubble interaction data feeds back into the model training. This loop helps the ranking system better understand which friends and content combinations resonate with users.
We’ve enhanced our existing video ranking formula, which includes several optimization goals, with a friend bubble ranking goal designed to maximize overall video engagement. It takes into account interaction metrics such as watch time, comments, and likes, and uses conditional probability terms. P(Video Engagement | Bubble Impression)predicting the likelihood that a user will engage with a video after viewing a friend bubble.
This is balanced by adjustable weights that manage the trade-off between social interaction and video engagement, allowing you to optimize social connections (making it easier for friends to find videos they like) and content quality. This double optimization captures the core value proposition of the friend content ranking system. That means connecting easily with passive friend discovery, providing entertainment with relevant content, and strengthening relationships by turning shared videos into natural touchpoints for conversation.
Client Infrastructure Behind the Scenes: Performance at Reel Scale
Reels are a performance-sensitive surface, so adding new per-video metadata is not as simple as adding another field. Additional work while scrolling or delayed playback can detract from the core user experience. When we integrated friend bubbles, we treated three constraints as non-negotiable.
- smooth scrolling
- No load latency regression
- Low CPU overhead for fetching and processing metadata
Facebook’s video distribution system already performs significant prefetching work prior to playback. Preload metadata, thumbnails, and buffered content before the video reaches the viewport. Locking friendbubble metadata retrieval to the same prefetch window provided several benefits. You can reuse cached results for stable data, avoid redundant CPU work, and limit wasteful network requests by using already optimized fetch paths.
Because the bubble data arrived along with the video content, the bubbles could be rendered at the same time as the video itself, eliminating the need for UI updates or redraws during playback.
I also made the animation strictly conditional. While interacting with active scrolling, animations are disabled to keep scrolling responsive. On low-end devices, even idle animations can degrade performance, so turn it off completely. In addition to additional optimizations to the underlying methods, this approach allows us to ship friend bubbles while maintaining core reel performance.
Why metadata needs to earn its place
A cleaner user interface is usually better, but it can backfire if new metadata adds noise or slows down the experience. Friend bubbles work because the signals are valuable. This adds meaningful social context to help people decide what’s worth watching.
By setting a conservative threshold at which friends are eligible to appear, we ensure that bubbles only appear when the relationship signal determined by the intimacy model between users is strong. This approach reduces clutter and improves the overall viewing experience, which is reflected in increased video playback time.
Impact and future of friend bubbles
Friend bubbles improve the relevance of your content and the quality of engagement. In user feedback studies, videos annotated with bubbles consistently receive higher interest scores and positive sentiment ratings than videos without bubbles.
Beyond relevance, bubbles increase not only the quantity but also the quality of app sessions. Users who view bubbles spend more time actively viewing and interacting with content, increasing their focus on longer sessions rather than short check-ins. This improvement primarily comes from deeper use of video. Bubble-related signals show delayed effects on long-term engagement patterns, suggesting that repeated exposure to content interacted with by friends leads to the formation of sustained interest over time.
By displaying content that their friends have engaged with, Bubbles can expose users to a wider range of topics and creators than they would encounter organically. Rather than just passively scrolling through this content, users actively engage with it through likes, comments, shares, and follows, showing that content recommended by friends can resonate even if it falls outside of their normal interests.
Not all friend signals are equal. Bubbles triggered by expressive responses, such as love or laughter, drive stronger downstream engagement than simple likes, especially in the case of comments and private shares, suggesting that expressive responses have stronger resonance. Also, engagement consistently increases with the number of friend bubbles displayed. This means videos that have interactions with multiple friends tend to perform better.
We are then extending the system to increase impact and robustness by extending friend-driven recommendations to additional surfaces and inventories while maintaining quality, improving cold start for users with limited friend graphs, and refining rankings and feedback signals to improve personalization.
Ultimately, this architecture shows how machine learning can enhance human connections at scale, making it easier for people to discover common interests and start conversations with the people who matter most. When your friends are enjoying something great, you can discover it too and talk about it together with just a tap.
For more information about Facebook bubbles, please visit: meta newsroom.
