Facebook Engineering


Meta engineering team explains the technical architecture behind Friend Bubbles on Facebook Reels, a feature that surfaces videos your friends have liked or reacted to. The system uses two ML models to estimate viewer-friend closeness: one trained on user surveys and one on platform interaction signals. Video relevance is handled by expanding the retrieval funnel with friend-interacted content and integrating social signals into multi-task ranking models via a continuous feedback loop. A conditional probability term P(video engagement | bubble impression) is used in the ranking formula. On the client side, bubble metadata is fetched within the existing video prefetch window to avoid scroll latency or CPU overhead. Results show bubble-annotated videos receive higher interest scores, longer watch sessions, and broader content discovery, with expressive reactions driving stronger downstream engagement than simple likes.

Friend Bubbles: Enhancing Social Discovery on Facebook Reels

An Overview of the Friend Bubbles System Architecture

Viewer-Friend Closeness: Identifying Friends With User-User Closeness Models

Video Relevance: Making the Ranking System Friend-Content Aware