Applying Embedding-Based Retrieval to Airbnb Search

TL;DR

This work addresses scaling Airbnb search to millions of listings per query by introducing an Embedding-Based Retrieval (EBR) system that represents both queries and listings in a shared embedding space. It proposes a two-tower model with offline per-listing embeddings and an online query encoder, coupled with a third-stage setwise re-ranker to produce top candidates for re-ranking. Key contributions include trip-based training data sampling to reflect multi-stage user journeys, an offline traffic-replay framework for realistic evaluation, and an online deployment using Inverted File Indexing (IVF) to balance recall and latency across product surfaces like Flexible Date Search and promotional emails. The results show significant offline recall improvements and online conversion gains (notably a ~0.31% relative uplift) along with substantial compute savings, demonstrating the practical viability and business value of EBR in a large two-sided marketplace.

Abstract

The goal of Airbnb search is to match guests with the ideal accommodation that fits their travel needs. This is a challenging problem, as popular search locations can have around a hundred thousand available homes, and guests themselves have a wide variety of preferences. Furthermore, the launch of new product features, such as \textit{flexible date search,} significantly increased the number of eligible homes per search query. As such, there is a need for a sophisticated retrieval system which can provide high-quality candidates with low latency in a way that integrates with the overall ranking stack. This paper details our journey to build an efficient and high-quality retrieval system for Airbnb search. We describe the key unique challenges we encountered when implementing an Embedding-Based Retrieval (EBR) system for a two sided marketplace like Airbnb -- such as the dynamic nature of the inventory, a lengthy user funnel with multiple stages, and a variety of product surfaces. We cover unique insights when modeling the retrieval problem, how to build robust evaluation systems, and design choices for online serving. The EBR system was launched to production and powers several use-cases such as regular search, flexible date and promotional emails for marketing campaigns. The system demonstrated statistically-significant improvements in key metrics, such as booking conversion, via A/B testing.

Figures (9)

• Figure 1: Core ranking models in the Airbnb search ranking stack, each phase scoring an order of magnitude less results than the previous one
• Figure 2: Example of search-based sampling to construct training data. The user makes a booking after doing five searches but their early searches without a booked listing (i.e. Search #1 and Search #2) are discarded
• Figure 3: Sampling when done at trip level as used in retrieval training data. Note that early searches in a user's journey are now retained and used to construct contrastive pairs
• Figure 4: Training data pipeline for the trip-based sampling with harder negatives as used in retrieval model training
• Figure 5: Two-tower architecture as used in EBR model. Note that the listing tower is computed offline daily