Real-time Personalization Using Embeddings for Search Ranking at Airbnb (KDD 2018)
Authors / Venue: Hristo Grigoriff and colleagues at Airbnb; KDD 2018 (Applied Data Science track) Source: https://dl.acm.org/doi/abs/10.1145/3219819.3219885
Key points
- Airbnb trained listing embeddings from booking-session data using a word2vec-style skip-gram objective, where listings clicked in the same session form a "sentence" and consecutive listings form contextual pairs.
- The training objective was adapted to the marketplace: booked listings are treated as a global context (final positive example), and listings from the same market (geographic region) are used as negative samples, ensuring learned similarity reflects booking intent rather than mere co-clicks.
- Listing embeddings support similar-listing recommendations and serve as real-time features in search ranking: similarity between the candidate listing and the user's most recently clicked listing (plus skipped-vs-clicked and price-difference signals) is fed into a GBDT ranker.
- User-type / listing-type embeddings are trained on cross-session sequences of user actions (segmented into user types by market, device, and behavior), capturing long-term interest personalization and enabling matching of users to listings even with no direct co-occurrence, including via candidate pre-ranking within each market.
- Cold-start listings are handled by locating nearest neighbors within the same geographic area and price bucket based on attributes rather than interaction data.
- Offline evaluation used perceived-similarity tests, A/B placement of similar listings, and downstream training-data utility; online A/B tests showed statistically significant improvements in booking conversion (NDCG gain of roughly 0.7% for the embedding features).
- The paper also documents practical production concerns: real-time feature computation latency, embedding retraining cadence (daily), and the effect of similarity features on ranking calibration.
- Behavior modeling for e-commerce search (behavior-driven query similarity with pre-trained language models)
- Query–product relevance labeling with large language models
- Deep learning for ads conversion prediction
- Original paper: *Real-time Personalization Using Embeddings for Search Ranking at Airbnb*, KDD 2018 — https://dl.acm.org/doi/abs/10.1145/3219819.3219885
Why it matters
This is one of the most cited industrial case studies of adapting negative-sampling embedding learning (word2vec lineage) to search and recommendation, showing how to:
1. encode marketplace-specific business signals (bookings, market locality) into the embedding objective; 2. consume embeddings both as retrieval/similarity structures and as real-time ranking features; 3. handle sparse interactions, cold-start items, and cross-session user modeling in production.