Problem
A short-video feed can show videos that the viewer's friends liked or reacted to. These videos sometimes rank low, not because they are weak but because the ranking model lacks user-user closeness context. A friend's reaction matters only if the friend is close to the viewer, and closeness is not observed directly. Any new label on a video must also not slow scrolling or loading.
Example from the article: in Facebook Reels, "friend bubbles" show which friends reacted to a video, and a tap on a bubble opens a one-on-one chat with that friend.
Solution in brief
Train two closeness models, one on survey answers and one on on-platform interactions, and use them to pick which friends count. Retrieve videos that close friends interacted with, add bubble features and a new task to the early and late multi-task ranking models, and blend a bubble objective into the ranking formula. Fetch bubble metadata in the existing client prefetch window.
Questions this case answers
- How to rank items that connections of the viewer engaged with when the value of that signal depends on relationship strength, which is not directly observed
- How to add a new per-item annotation to a latency-sensitive feed without slowing scrolling or loading
- How to rank videos that friends interacted with so that the closest friends' activity counts most?
- How to estimate how close two users are when there is no explicit closeness signal?
Tools and methods
- Survey-based closeness model — close / not-close classifier trained on survey labels.
- Interaction-based closeness model — trained on likes, comments and reshares seen with bubbles.
- Friend-aware retrieval — candidates from top close friends' interactions.
- Multi-task, multi-label (MTML) ranking models — early and late stage, with bubble features and a bubble task.
- Bubble ranking objective — weighted blend of social interaction and video engagement.
Google or a one-time email code. No password.