Friend Bubbles: Enhancing Social Discovery in Facebook Reels

Friend Bubbles shows videos that friends have liked or reacted to as bubbles in Facebook Reels, making content discovery and starting conversations easier. With a single tap, users can immediately open a 1:1 conversation with the friend who reacted to that Reel.

The system uses viewer-friend closeness and video relevance together to prioritize content where the friend relationship is strong and the context with the video also matches. Videos with multiple accumulated friend reactions are interpreted as a stronger signal of shared interest, and this social signal creates a feedback loop that further improves recommendation quality.

Friend-to-friend intimacy is judged using two models.

• Survey-based closeness model: trained on social graph/user attributes such as mutual friends, interaction patterns, user-provided location, number of friends, and number of posts.
• On-platform activity-based model: predicts contextual closeness within the platform using actual interaction signals such as likes, comments, and reshares at the time the bubble is displayed.

The survey-based model creates labels through a lightweight binary survey asking a random sample of users "are you close to this connection in real life," and is retrained periodically. In production, it performs inference weekly on trillions of person-to-person connections.

On the video recommendation side, two things are key. First, at the retrieval stage, videos that close friends identified by the closeness model have reacted to are elevated as candidates, widening the top of funnel. Then at the ranking stage, friend-bubble interactions are incorporated as features and into multi-task, multi-label (MTML) tasks, so the model better learns relationship strength and social meaning.

In particular, the reason existing models rated friend content low is seen not as a matter of low quality, but because they lacked user-user closeness context. To address this, a continuous feedback loop was built that feeds friend-bubble interaction data back into training, and a friend-bubble ranking objective was introduced that includes P(video engagement | bubble impression), the conditional probability of video engagement given a bubble impression.

On the client side, strong constraints were imposed to avoid harming Reels' performance.

• smooth scrolling
• no regressions in load latency
• low CPU overhead for metadata fetch and processing

Since the video delivery system already prefetched metadata, thumbnails, and buffered content before playback, friend-bubble metadata retrieval was attached to the same prefetch window. This enabled cache reuse, reduced unnecessary CPU work, and minimized redundant network requests, allowing bubble data to arrive simultaneously with the video and render without intermediate redraws.

Animations are turned off during scrolling and interaction to preserve responsiveness, and on low-end devices they are disabled entirely even in the idle state. Additionally, a conservative threshold was applied so that too many bubbles wouldn't clutter the screen, ensuring only friends with strong relationship signals are shown.

The effect was clear. Videos with bubbles attached received higher interest scores and more positive sentiment in user surveys, and session length increased more in longer viewing sessions than in short check-ins. In particular, expressive reactions like love or laughter produced stronger downstream engagement in comments and private shares than simple likes.

Going forward, the plan is to expand friend-driven recommendation to more surfaces and inventory, improve cold start for users with small friend graphs, and further refine ranking and feedback signals. Ultimately, this structure demonstrates that recommendation systems can enable the discovery of shared interests even at massive scale, and can turn those interests into actual conversations.