Problem
Many travel-platform users are still exploring: they have no fixed destination, dates or preferences. They visit less often, are less likely to book soon, and search broad areas such as "France". Two groups need different handling: active users who searched recently, and dormant users who booked before and are now planning without having opened the app. The goal is to spark ideas and narrow the choice, lifting engagement and bookings.
Example from the article: user A searched the California Bay Area last week and wants cheaper listings there; user B booked several times in 2025, has not returned, and is thinking about a summer 2026 trip.
Solution in brief
Treat each user action (booking, view, search) as a token and predict the next destination with a transformer, as a language model predicts the next word. For every booking, build 14 training examples: 7 mimic active users and 7 mimic dormant users. Add region-level and city-level prediction heads so the model learns how cities group into regions. The output feeds search autosuggest and abandoned-search emails.
Questions this case answers
- How to predict what a user wants next when their intent is still vague and their last activity may be recent or months old
- How to build training examples from past conversions so that one model serves both users who are close to converting and users who are dormant
- How to make a model that predicts a leaf category also use the hierarchy that groups those categories
- How to recommend destinations to users who have not yet picked where, when or what to book?
Tools and methods
- Transformer over user action sequences — encodes bookings, views and searches as one token sequence and predicts the destination
- Time-based training examples — 7 active and 7 dormant examples per booking
- Multi-task learning with region and city heads — injects geographic hierarchy into the representation
- Online A/B test — measures the booking impact of autosuggest
Google or a one-time email code. No password.