Recommending Travel Destinations to Help Users Explore

Airbnb describes how they built a destination recommendation model to help exploratory users in the trip planning stage discover and narrow down travel destinations.

• •A transformer-based model treats each user action (booking, view, search) as a token, encoding city, region, and days-to-today embeddings to capture both short-term and long-term interests.
• •Training data is designed with 14 examples per booking: 7 for active users (1–7 days before booking using full history) and 7 for dormant users (8–365 days before, using only booking history).
• •Multi-task learning with multiple prediction heads at the final layer jointly trains region-level and city-level destination prediction to improve geolocation representations.
• •The model is deployed in two features: autosuggest (city recommendations shown when users click the search bar) and abandoned search email notifications.