Personalizing Airbnb search by learning from the guest journey
Researchers at Airbnb developed a Transformer-based sequence model to personalize search results by learning from guest behavior. The model splits the guest sequence into two parts: a long-term sequence of infrequent but informative events from the past seven years, and a short-term sequence of recent listing views. This allows the model to capture both historical booking patterns and immediate browsing behavior, improving ranking by 1.48%. To address computational and efficiency challenges, the team employed several strategies, including batching of searches, bucketization of sequences by length, and sparse calculation of searches. These measures reduced model training costs and improved throughput by around 4x. A decoupled serving design was also implemented, separating encoder and ranking model stages to minimize latency and efficiently store guest embeddings. The new ranking system was rolled out in three stages, each tested through A/B testing to rigorously evaluate key business and safety metrics. The final system demonstrated a 1.48% improvement in ranking relevance compared to the existing system, with no