Personalizing Airbnb search by learning from the guest journey

Summary

Airbnb replaced hundreds of hand-crafted ranking features with a Transformer-based sequence model that captures both long-term booking history and short-term browsing behavior. To keep latency low and throughput high, sequence representations are generated via daily batch jobs and combined with real-time queries using a co-trained setwise ranker.

Takeaways

1. 1

   Guest histories are split into long-term events (up to 7 years of bookings and cancellations, capped at 80 events) and short-term events (past 21 days of listing views, capped at 200 events).

2. 2

   Training throughput improved fourfold by using causally masked search batching to share a single forward pass across multiple searches, sequence-length bucketizing, and sparse calculations.

3. 3

   Serving latency is minimized by precomputing sequence embeddings in daily batch jobs and retrieving them in real time for a setwise ranker that scores listing candidate sets together.