How to Recommend Destinations to Help Users in the Exploration Stage

Airbnb built a destination recommendation model for users in the exploration stage who haven't yet decided on a travel destination. Since the destination is a key variable that shapes accommodations, itinerary, and budget, recommending destinations alone can reduce decision-making friction and increase booking conversion.

This model borrows ideas from language modeling, treating user behavior like tokens. It feeds in sequences of booking, view, and search history, where each action is represented as the sum of city, region, and days to today embeddings, and also incorporates contextual information such as the current point in time, allowing it to simultaneously learn both short-term and long-term interests.

In designing the training data, the team handled active users and dormant users separately. For each booking, they created a total of 14 training examples, 7 of which use the most recent booking/view/search data from 1 to 7 days before the booking to simulate the late booking stage.

The remaining 7 examples are dormant user examples randomly sampled from 8 to 365 days before the booking. In this case, only booking data is used, reflecting exploration patterns from the early planning stage where a specific destination hasn't yet been decided.

To improve location understanding, multi-task learning was also applied. Region-level and city-level prediction heads were each placed in the final layer, and training both levels of prediction together produced a richer geolocation representation.

This model was applied to two features.

• Autosuggest: Shows multiple city recommendations when the search bar is clicked.
• Abandoned search email notifications: Sends emails containing accommodations in alternative regions predicted by the model to users who abandoned their search.

In online A/B tests, significant booking improvements were confirmed particularly in regions where English is not the primary language. Additionally, the recommendations helped not only users who hadn't yet decided on a destination, but also users who had room to move to nearby, cheaper cities.

Ultimately, this framework provides a foundation that can be extended to broader personalization across travel planning as a whole. The authors believe the same approach can later be applied to predicting other preferences, such as travel timing and price preferences.