Airbnb's Proximity Features: Personalizing for Cold-Start Users
How Airbnb uses geographic signals to personalize without individual user data.
1The cold-start challenge
Airbnb faces a significant personalization problem with new or logged-out users who lack search history or identifiers. Traditional ML models rely on individual user data, which isn't available for approximately 30% of traffic from ads and organic search. Without these signals, users receive generic recommendations that don't reflect local preferences or travel patterns. The challenge intensified with GDPR and browser restrictions limiting tracking capabilities. This created a gap where personalization was weakest precisely where it mattered most - during first impressions that convert visitors.
2Geographic correlation insight
The solution leverages a fundamental observation: users from the same area exhibit similar travel behaviors. Someone in Seoul is more likely to book Jeju Island or Osaka than a New Yorker would. Airbnb's system groups users by coarse IP-derived locations into 'proximity buckets' of about 1,000 people. These buckets aggregate short-term searches, long-term booking patterns, and local metadata. The approach requires no login or persistent identifiers, working purely through geographic clustering. This maintains privacy while solving the cold-start problem through crowd-sourced behavioral signals.
3Adaptive clustering algorithm
The technical implementation handles extreme variations in user density across locations. A two-phase algorithm creates fine-grained tiles for cities (using IP hash buckets) and coarser groupings for rural areas. This ensures each bucket contains approximately 1,000 users whether in Tokyo or a sparsely populated region. The system runs efficiently at global scale, with stable clusters that haven't required reconfiguration since 2023 deployment. At serving time, IP resolution to proximity keys happens in real-time as a non-blocking operation, falling back gracefully when needed.
4Privacy-first design
The system embeds privacy protections at multiple levels. Features reflect group behavior (~1,000 users) never individuals. Geo-IP coordinates are intentionally coarse, centered on population areas rather than specific addresses. The pipeline integrates with consent management systems, excluding users who haven't opted in. Data governance controls ensure compliance with regulations like GDPR. This design allows personalization while avoiding the need for persistent identifiers or detailed location tracking that might concern privacy-conscious users.
5Marketing landing page impact
Proximity features showed measurable improvements where cold-start rates were highest - paid advertising and organic search landing pages. Previously, these pages showed static listings due to sparse feature vectors. With proximity signals, the recommender could surface properties aligned with local booking patterns. The change brought meaningful personalization to traffic segments that previously received none, improving conversion rates without compromising privacy standards or requiring additional user data.
6Location-aware suggestions
The homepage AutoSuggest feature demonstrated particularly clear benefits. For users without history, it previously displayed generic global destinations (Paris, London). With proximity features, suggestions became location-relevant - a Beijing user might see Hong Kong and Tokyo instead. Experiments showed gains concentrated among never-booked and dormant users. The treatment arm consistently displayed previously underrepresented destinations like Kuala Lumpur and Dubai to users in relevant regions, increasing suggestion diversity.
7Architecture advantages
The system design offers multiple technical benefits. Proximity keys function analogously to user IDs, allowing any model using personalization to serve cold-start users. Features update daily based on aggregated group behavior. The implementation requires no persistent storage of individual user data. Lookups are soft dependencies that don't block core functionality. This architecture makes the solution scalable across Airbnb's global traffic while maintaining performance standards for real-time personalization.
8Future implications
Airbnb's approach demonstrates that advanced personalization doesn't require compromising user privacy. The success with proximity features suggests broader applications for zero-party data strategies. The team is extending the framework to engagement emails using coarse location signals. The reusable design pattern could influence how other platforms handle cold-start scenarios amid increasing privacy regulations. As tracking limitations grow, geographic and group-based signals may become crucial alternatives to individual behavioral data.