Personalizing Airbnb search by learning from the guest journey

2026-07-21 · Source: The Airbnb Tech Blog - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Advanced, medium

Summary

Airbnb has developed a Transformer-based sequence model to personalize search results by learning from the full guest journey. This new system replaces an older approach reliant on hundreds of hand-crafted features, which became difficult to scale and limited in expressiveness. The model processes two distinct event sequences: a long-term sequence of up to 80 infrequent events like bookings and reviews from the past seven years, and a short-term sequence of up to 200 listing views from the last 21 days. To ensure efficiency, Airbnb implemented strategies such as batching searches, which improved training throughput by approximately 4x, and decoupled inference for low serving latency. Rigorous A/B testing confirmed significant improvements, with the long-term sequence alone boosting uncanceled bookers by +0.31% and the combined system, including a setwise ranker, achieving a total offline NDCG improvement of 3.78%. The model also generalized effectively to promotional email ranking, increasing email clicks by +5.04%.

Key takeaway

For AI Engineers building personalized recommendation or search systems, you should re-evaluate reliance on hand-crafted features. This work demonstrates that Transformer-based sequence models, by encoding your users' full journey over years, can significantly outperform traditional methods. Implement a dual-sequence approach for long-term history and recent activity, and prioritize efficiency strategies like search batching and decoupled inference to ensure your system scales and maintains low latency. Your investment in sequence modeling will yield substantial improvements in user engagement and conversion.

Key insights

Transformer-based sequence models enhance search personalization by encoding comprehensive guest journey data.

Principles

• Guest behavior is a rich sequence signal.
• Decouple inference for real-time serving.
• Rigorous A/B testing validates model impact.

Method

Split guest event sequences into long-term (7 years, 80 events) and short-term (21 days, 200 views) for Transformer encoding, optimizing with search batching and decoupled inference.

In practice

• Adopt sequence models for user preference learning.
• Optimize training with search batching.
• Decouple sequence encoding from real-time ranking.

Topics

• Transformer Models
• Sequence Modeling
• Search Personalization
• Recommendation Systems
• Machine Learning Efficiency
• A/B Testing

Best for: Research Scientist, Machine Learning Engineer, AI Engineer, AI Scientist

Related on AIssential

• JourneyFormer: Encoding Airbnb Guest Journey with Sequence Modeling — Successfully deploying sequence models in production requires careful design decisions for data, architecture, and optimization.
• Sparse Personalized Text Generation with Multi-Trajectory Reasoning — PaT enhances cold-start LLM personalization by integrating style and topic cues via iterative, differential-reward-based reasoning.
• Proximity Features: Privacy-Compliant Cold-Start Personalization at Airbnb — Proximity Features enables privacy-compliant cold-start personalization by aggregating geo-IP data for user groups, boosting bookings for new users.
• Bi-NAS: Towards Effective and Personalized Explanation for Recommender Systems via Bi-Level Neural Architecture Search — The Bi-NAS framework optimizes recommender system explanations by combining bi-level neural architecture search with LLM-enhanced, personalized justification generation.
• From Clicks to Conversions: Architecting Shopping Conversion Candidate Generation at Pinterest — Optimizing for sparse offsite conversions requires multi-task learning, robust data design, and parallel architectural components.
• From RNNs to Transformers: How Sequence Models Evolved — Sequence models evolved by overcoming sequential processing and fixed context bottlenecks through dynamic attention and parallel computation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Airbnb Tech Blog - Medium.