Problem
A large share of Airbnb visitors arrive logged out, without recent searches or bookings, especially from paid ads and organic search. Models for ranking, destination suggestions and landing pages key their features on a user identifier, so these visitors get almost no personalization. Privacy rules (GDPR) and the loss of third-party cookies make persistent identifiers unusable for this traffic. Even logged-in users sometimes have stale or sparse features.
Example from the article: on the Homepage, users with no history saw a static list (Paris, Barcelona, London, Rome) wherever they browsed from.
Solution in brief
Replace the user id with a proximity key: a geographic bucket of about 1,000 users derived from the visitor's IP. Features are the aggregated searches and bookings of the whole bucket, computed daily. Any model keyed on a user id can be keyed on the proximity key instead.
Questions this case answers
- How to personalize model predictions for new or anonymous users when no persistent user identifier is allowed or available
- How to build group-level features from coarse context signals so that a model keyed on a user id can be reused for users without an id
- How to personalize recommendations for logged-out or first-time visitors who have no history and no persistent identifier?
- How to build features that work under GDPR and cookie restrictions without tracking individual users?
Tools and methods
- Proximity key — compact group key: a quantized latitude/longitude tile plus, in dense areas, an IP-hash bucket index.
- Adaptive two-phase clustering — builds buckets of about 1,000 users, fine in cities and coarse in rural areas.
- Daily bucket features — short-term engagement, long-term bookings, bucket metadata.
- Key-value serving with a soft dependency — features are fetched by key; on timeout the model runs without them.
Google or a one-time email code. No password.