Personalizing Airbnb search by learning from the guest journey

Daochen Zha21 Jul 2026

Problem

Airbnb search ranking captured guest preferences through hand-crafted aggregates such as total past bookings or average listing price. With hundreds of such features, the set became hard to scale and limited in expressiveness. Guest histories are long, views dominate them (some guests have hundreds of thousands), and bookings are rare and deliberate. Training on hundreds of millions of search-label pairs is also expensive.

Example from the article: a guest planning a trip to San Francisco browses dozens of listings over several days, on top of many earlier bookings.

Solution in brief

Replace aggregate features with a Transformer encoder over the guest's event sequence, split into a long-term part (rare, informative events) and a short-term part (recent views). The encoder runs as a daily batch job. The ranker reads the stored embedding at query time. Three training optimizations give about 4x throughput.

Questions this case answers
  • How to learn user preference representations from long, noisy event histories where the target action is rare and most events are low-intent browsing
  • How to train a sequence encoder on hundreds of millions of ranking examples and serve it within a tight query-time latency budget
  • How to personalize search ranking using a guest's multi-year event history instead of hundreds of hand-built aggregate features?
  • How to model very long user event sequences dominated by views when the target (a booking) is rare?

Tools and methods

  • Transformer sequence encoder — turns a guest's event history into one embedding used as a ranking feature
  • Long-term and short-term sequences — keep rare informative events separate from the flood of views
  • Unified embedding table — one hashed table for high-cardinality IDs (listing, host, geography)
  • Setwise ranker — scores a candidate set jointly, co-trained with the encoder
  • Batching of searches, length bucketing, sparse search calculation — training speed-ups
  • Daily batch inference — keeps query latency low
Continue reading — it's free

Google or a one-time email code. No password.