A different feed for everyone
Two users who follow the same accounts should still see different feeds, because they engage with different things. Personalization makes ranking depend on the individual viewer, not just the post.
Where the personal signal comes from
- A user profile captures interests inferred from past behavior, what they like, watch, and linger on.
- Interaction history weights authors and topics the user engages with most.
- Context such as time of day and device can shift what is surfaced.
The ranking model takes both post features and these user features, predicting engagement for this specific viewer.
The exploration problem
Pure relevance creates a filter bubble. If the feed only shows what a user already likes, it never learns new interests and the user gets bored. So systems deliberately explore, mixing in some uncertain content to gather signal.
This is an explore versus exploit tradeoff. Exploit shows known winners, explore tests new things to keep learning.
Feedback loop risk
Personalization learns from clicks, but clicks are shaped by what was shown. Without care the loop narrows over time. Diversity rules and exploration counter that drift.
Key idea
Personalization ranks posts using user features from behavior alongside post features, and deliberately explores new content to avoid filter bubbles and narrowing feedback loops.