Airbnb
Search Context
Airbnb runs two distinct search products:
- Listing search: finding homes and apartments to rent (the core product)
- Experiences search: finding local tours, classes, and activities hosted by individuals
Both are marketplace ranking problems — the goal is not just relevance but probability of booking. Airbnb has published extensively on ML-based ranking and embedding-based personalization.
Search Challenges
- Conversion signal as relevance: bookings (not just clicks) are the true label — but sparse
- Cold start for new listings/experiences: no historical engagement data
- Personalization at scale: pre-computing rankings for 1M+ users daily
- Diversity vs. precision tradeoff: showing only the single most relevant experience category hurts discovery
- Multi-objective: quality, popularity, diversity, and business goals must be balanced in one model
Tech Stack
- Gradient Boosted Decision Trees (GBDT) for core ranking
- Category Intensity features derived from user click history with recency decay
- Offline pre-computation of personalized rankings + online re-ranking for query-time features
- 32-dimensional listing embeddings trained on 800M booking sessions (word2vec-style)
Use Cases
Published Articles
Compare With
- Booking.com — the closest published counterpart. Both are accommodation marketplaces that treat booking rather than clicks as the true label, and both project listings into a latent space with word2vec-style training over user action sequences instead of relying on human-authored attributes. See Beyond Algorithms - Ranking at Scale at Booking.com.
- Skyscanner — travel meta-search with purchase-completion labels