Skyscanner
Global travel meta-search engine. Search challenge: ranking flight itineraries by predicted relevance to booking intent, not just by price.
Search Context
Traditional flight search sorts by price. Skyscanner applied Learning to Rank to surface flights that better match user booking intent — shifting from a commodity sort to a relevance-ranked experience.
Key Engineering Work
- LTR model: logistic regression on user search history, behavioral signals, flight attributes
- Training labels: purchase completion (not just clicks) — aligns model to business objective
- Offline evaluation: MAP and MRR predicted online conversion improvement
- Result: ML ranking drove more purchases into the recommendation widget vs. rule-based variant
Key Lesson
Purchase completion is a better relevance signal than clicks. An LTR model optimized for booking outperforms rule-based flight ranking even with a relatively simple logistic regression model.
Articles
- Learning to Rank for Flight Itinerary Search — full LTR workflow, feature engineering, A/B results