Facebook Reels Introduces Friend Bubbles
Friend bubbles are now live on Facebook Reels, highlighting content your friends have liked or reacted to. This feature enhances the social experience by helping users discover new content and connect over shared interests.
The system employs two complementary machine learning models to identify closeness between users. The survey-based closeness model utilizes a wide array of signals, including social graph features and user attributes, to build a comprehensive understanding of relationships. This model is trained using a binary survey that asks users to predict if they feel close to a specific connection. In production, this model performs weekly inference across trillions of person-to-person connections.
Technical Architecture of Friend Bubbles
Friend bubbles also incorporate a context-specific closeness prediction model. This model is trained on on-platform activity signals, capturing how users interact with each other in real-time. This allows the system to understand how likely a viewer is to value content recommended by a friend, based on their existing interactions.
The recommendation system integrates friend-bubble interaction signals into both early-stage and late-stage ranking multi-task, multi-label (MTML) models. This continuous feedback loop ensures that friend-bubble interaction data informs model training, allowing the system to better understand which friend-content combinations resonate with users.
Ranking and Performance Optimizations
Two key strategies are employed to ensure high-quality, friend-interacted content reaches users. The retrieval stage sources candidate videos based on close friends, expanding the top of the recommendation funnel. This ensures that valuable friend content is considered.
Facebook’s video delivery system already prefetches significant content. By pinning friend-bubble metadata retrieval to this prefetch window, the system reuses cached results, avoids redundant CPU work, and limits wasted network requests. Because bubble data arrives with video content, bubbles render simultaneously with videos, preserving core Reels performance with disabled animation during active scrolling to maintain responsiveness.
Impact and Future Development
Friend bubbles demonstrably improve content relevance and engagement quality. User feedback surveys indicate that bubble-annotated videos consistently receive higher interest scores and more positive sentiment. Users interacting with bubbles spend more time actively watching and engaging with content.
Expressive reactions like love or laughter on bubbles drive stronger downstream engagement compared to simple likes. Engagement also scales with the number of friend bubbles shown. Future plans include scaling the system to improve cold start for users with limited friend graphs and refining ranking signals for better personalization. For more information, visit the Meta Newsroom.
Tools We Use for Working with AI:
