Airbnb Engineers Develop Transformer-Based Model for Travel Destination Recommendations

Airbnb engineering teams have developed a sophisticated destination recommendation model designed to assist users in the exploratory phase of trip planning. Recognizing that many users lack clear itineraries and search broadly for inspiration, the team created a framework to predict destination intent by analyzing user actions on the platform. Inspired by language modeling, the system utilizes transformer architectures to process sequences of user behaviors, including booking, viewing, and search histories. The model effectively integrates diverse signals, balancing long-term interests with short-term contextual factors like seasonality and current time. A key challenge addressed is differentiating between active users, who exhibit recent engagement, and dormant users with sparse data. By treating each user action as a token and encoding rich geolocation knowledge, the model generates holistic predictions of travel intent. This technology powers practical applications such as autosuggest features and abandoned search email notifications, aiming to reduce decision friction, spark inspiration, and ultimately drive higher engagement and conversion rates for the travel platform.