Airbnb Introduces Proximity Features for Cold-Start Personalization Without User Identity

Airbnb has deployed Proximity Features, a new class of machine learning features designed to solve the cold-start problem for users without login history or persistent identifiers. By leveraging geographic correlation, the system aggregates collective signals from nearby users to personalize search ranking, recommendations, and marketing landing pages immediately upon arrival.

How Proximity Keys Work

The core mechanism is the proximity key, a compact group key representing a local geographic cluster of approximately 1,000 users. Derived from a user's IP address, this key functions as an aggregation analog to user_id, allowing models to condition on group behavior rather than individual history. The key encodes a quantized latitude/longitude tile, with additional IP hash bucket indices in densely populated areas to maintain granularity.

Features are computed daily across three categories:

• Short-term engagement: Top destinations, room types, and median prices from recent searches.
• Long-term booking patterns: Booked destinations and travel party signals.
• Aggregate bucket metadata: Bucket size and geographic density.

Adaptive Clustering and Stability

To handle varying geographic densities, Airbnb uses a two-phase adaptive clustering algorithm. In dense areas like city centers, coordinates are subdivided using IP hash buckets. In sparse rural areas, multi-pass coarsening widens geographic tiles until each bucket accumulates roughly 1,000 users. This approach ensures stable buckets that have remained valid in production since 2023 without re-clustering, relying on daily refreshes for new IPs.

Privacy and Production Results

Privacy is integrated at every layer: features reflect group aggregates rather than individuals, clustering excludes non-consenting users, and geo-IP coordinates are coarse. The system operates as a soft dependency, meaning personalization lookups do not block core request paths if they time out.

Production A/B experiments demonstrated measurable lifts:

• Marketing landing pages: Enabled meaningful personalization for cold-start traffic previously served static listing cards.
• Homepage AutoSuggest: Replaced generic global lists with location-aware suggestions for new users, showing directional gains for never-booked and dormant users.
• Engagement emails: Currently undergoing experiment validation for personalized campaigns using coarse location signals.

The full technical details are available in a paper accepted to the TSMO Workshop at KDD 2026.