Airbnb describes a new search system that learns from a guest’s entire history, with the goal of displaying more relevant listings when they are needed. In the company’s technical post, the model is based on a Transformer and uses both long-term travel history and short-term viewing activity.
The approach replaces, according to Airbnb, the exclusive reliance on hundreds of manually designed features, which had become increasingly difficult to maintain and scale. Instead of static descriptions, the system attempts to build a more complete picture of each guest’s preferences.
How the model is structured
The company divides the guest’s history into two parts. The first, the long-term sequence, includes important events from the past seven years, such as bookings, reviews and cancellations, with a maximum of 80 events.
The second, the short-term sequence, collects the listings the guest has viewed over the past 21 days and is limited to 200 events. According to Airbnb, these limits have been designed so that only the longest data sequences, representing approximately 2% of the total, need to be truncated.
In this way, the system combines long-term history with recent behavior. This allows it to take into account not only what a guest has chosen in the past, but also what they appear to be looking for at that moment.
To process this information, the model uses a shared data system for elements such as listings, hosts and geographic areas.
Where the optimization focuses
Airbnb says that a significant part of the improvement came from the way searches are organized. Instead of repeating the same processing from scratch for every new search, the system processes the guest’s history once and then uses the result for individual searches.
The company also used techniques that reduce unnecessary calculations, grouping data according to its size and avoiding the processing of data that does not add meaningful information. Overall, according to Airbnb, the three interventions deliver approximately a fourfold improvement in model training speed.
In day-to-day operation, the system regularly updates the guest’s profile with new data. When a new search is made, it combines this information with the details of the specific search and then ranks the available listings.
What the tests showed
Before wider deployment, Airbnb says it conducted tests lasting approximately three weeks and activated the system in three phases. First, it tested the long-term history, then added recent listing views and later introduced an additional model to improve the ranking of results.
In the initial tests, Airbnb found that using the guest’s long-term history improved the quality of search results by 0.44%. When recent listing views were added, the overall improvement reached 1.48%. The addition of another model for ranking results increased the overall improvement to 3.78%.
Tests with real users also showed positive results. The use of long-term history increased by 0.31% the number of guests who completed their booking without cancellation and by 0.38% views. With the addition of recent searches and views, the increases reached 0.55% in guests with a completed booking without cancellation, 0.82% in uncanceled nights and 0.90% in views.
The same technology was also used in Airbnb’s promotional emails to determine which listings would be shown to guests. According to the company, applying the system increased email clicks by 5.04%, while smaller increases were also recorded in completed bookings and uncanceled nights.
What this means for search on the platform
Airbnb’s main conclusion is that the system does not only look at what a guest has done recently, but also at what their overall history indicates about their current preferences. For hosts and property managers, this means that the order in which listings appear is becoming increasingly personalized. Two guests making similar searches do not necessarily see exactly the same results, as Airbnb can take each guest’s history and recent activity into account.
This makes factors such as listing quality, photos, reviews, price and the overall experience offered by a property even more important, as the system attempts to match each guest with the options it considers most likely to interest them.
The company also says that the project is not yet complete. Among the next areas it is exploring are more frequent updates to guest profiles, additional data such as wishlists and map interactions, as well as further improvements to the way listings are ranked.
Frequently Asked Questions
What problem is Airbnb trying to solve?
The company is trying to use the guest’s full history so that search can return more relevant listings based on both past bookings and recent behavior.
What does this mean for a host?
Airbnb is using more data to determine which properties are most likely to interest a particular guest. Therefore, personalization in search is becoming even more important for the visibility of listings.
