Why personalize
Two users typing the same query may want different things. Personalization adjusts ranking using signals about the user, like past clicks, location, and language.
Signal sources
- Long term profile from history, capturing durable interests.
- Short term session context, capturing the current task.
- Contextual signals like device and location.
Where it applies
Personalization usually acts as a reranking layer on top of a query independent base ranking. The base retrieval stays the same; personalization nudges the order. This keeps the heavy index work shared across users.
Risks to manage
- Filter bubbles where users only ever see agreeable content. Inject diversity to counter this.
- Privacy since profiles are sensitive; signals must be stored and used carefully.
- Cold start for new users with no history; fall back to popularity.
Diagram
Key idea
Personalization reranks shared base results using long term, session, and context signals, while guarding against filter bubbles, privacy risk, and cold start.