Bulles d'amis : améliorer la découverte sociale sur Facebook Reels

Friend bubbles in Facebook Reels highlight Reels your friends have liked or reacted to, helping you discover new content and making it easier to connect over shared interests. This article explains the technical architecture behind friend bubbles, including how machine learning estimates relationship strength and ranks content your friends have interacted with to create more opportunities for meaningful engagement and connection. Friend bubbles enhance the social experience on Facebook Reels by helping you discover content your friends enjoy, creating a shared viewing experience and sparking new conversations. With a quick tap on a bubble, you can start a one-on-one conversation with any friend who has engaged with that Reel. This feature combines social and interest signals to recommend more relevant, personalized content while making it easier to start conversations with the people who matter most to you. When videos connect to both personal interests and friend-related interests, they create a feedback loop that improves recommendations and strengthens social connections. An Overview of the Friend Bubbles System Architecture The friend bubbles recommendation system includes several components that work together to surface relevant, friend-interacted content by blending video-quality signals with social-graph signals: Viewer-Friend Closeness (Whose Interactions Matter Most): Identifies which friends’ interactions are most likely to interest the viewer. Video Relevance (What Videos to Show): Ranks videos that are contextually relevant to the viewer. Multiple friend interactions on the same video often signal stronger shared interest and higher relevance. Content surfaced through friend connections also tends to be high quality, creating a reinforcing loop: Social discovery increases engagement, and that engagement further strengthens the social graph. Viewer-Friend Closeness: Identifying Friends With User-User Closeness Models Friend bubbles rely on two complementary machine learning models to identify which connections a person feels closest to. One model is based on user survey feedback; the other is based on on-platform interactions. The survey-based closeness model draws on a broad set of signals, including social-graph features (mutual friends, connection strength, 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 in which a randomly selected group of Facebook users is asked whether they feel close to a specific connection in real life. The survey is structured as a close vs. not-close prediction problem, refreshed regularly to keep labels current, and includes questions that act as proxies for offline relationship strength (such as how often two people communicate). In production, the model runs weekly inference over trillions of person-to-person connections across Facebook friends. While survey-based closeness provides a strong foundation, friend bubbles also use a context-specific closeness prediction model trained on on-platform activity signals, using real interactions that occur when bubbles are shown (for example, likes, comments and reshares). This enables the model to capture closeness in context — how likely a viewer is to value content recommended by someone in their friend graph based on how they interact with each other on the platform. Our approach emphasizes connection quality over quantity. While bubble prevalence naturally rises with larger friend graphs, showing more bubble videos does not necessarily increase user engagement. The goal is to surface the right friend connections — those most likely to make the social context meaningful — using a combination of existing closeness signals and surface-specific features that better reflect the relationship dynamics behind friend-driven recommendations. Video Relevance: Making the Ranking System Friend-Content Aware We use two key strategies to ensure high-quality, friend-interacted content can move through the recommendation funnel and reach users: expanding the top of the funnel, and enabling models to rank friend-bubble content effectively through a continuous feedback loop. Sourcing Inventory: Expanding the Top of Funnel The retrieval stage sources candidate videos based on close friends, as identified by the closeness model described above. By explicitly retrieving friend-interacted content, we expand the top of the funnel to ensure sufficient candidate volume for downstream ranking stages. This is important because, without it, high-quality friend content may never enter the ranking pipeline in the first place. Enabling Models to Rank Friend Content Effectively Through a Continuous Feedback Loop A key insight from our development process was understanding why friend-interacted videos sometimes