Personalization without user identity | Airbnb Engineering & Data Science
AI Summary
Airbnb shares how they address the cold-start problem and privacy constraints for anonymous, logged-out, or first-time users by introducing Proximity Features. Instead of relying on individual user IDs, the system groups roughly 1,000 nearby users into geographic and IP-hashed clusters called proximity keys. By aggregating short-term engagement and long-term booking patterns within these local buckets, machine learning models can deliver location-aware recommendations without tracking individual history or violating privacy standards.
Key Takeaways
- Proximity keys group approximately 1,000 nearby users using a combination of quantized latitude/longitude tiles and IP hash buckets.
- A two-phase adaptive clustering algorithm dynamically adjusts resolution, providing fine-grained tiles in dense city centers and coarser groups in rural regions.
- Features are computed daily across short-term engagement, long-term bookings, and bucket metadata, and are served via a low-latency distributed key-value store with soft dependencies.
Why it matters
Working engineers facing strict privacy regulations (like GDPR) and cookie deprecation often struggle to personalize experiences for anonymous or cold-start users. This architecture demonstrates how to successfully leverage aggregate crowd signals via adaptive geographic clustering, bypassing individual user tracking entirely while improving conversion and relevance on marketing landing pages and homepages.
Conversation 0
Be the first to share your perspective.