Learning to Rank for Flight Itinerary Search
Source: https://hackernoon.com/learning-to-rank-for-flight-itinerary-search-8594761eb867 Author: Neal Lathia (Senior Data Scientist, Skyscanner London)
Summary
Case study of applying Learning to Rank to flight search at Skyscanner. Demonstrates the full LTR workflow from feature engineering through offline evaluation to A/B testing.
Problem
Traditional flight search sorts by price. LTR objective: rank flights by predicted relevance to the user’s actual booking intent, surfacing better matches above the standard list.
Technical Approach
Model: Logistic regression (initial experiments) Features: User search history, behavioral signals, flight attributes — joined, transformed, and reshaped into structured training data Labels: Binary relevance = purchase completion (not just clicks)
Evaluation
- Offline metrics: Mean Average Precision (MAP), MRR (Mean Reciprocal Rank)
- Online: A/B test comparing ML vs. rule-based vs. control
Results
ML ranking model drove more purchases into the recommendation widget than the rule-based variant. No significant difference in search effort (filtering/re-sorting frequency).
Lessons
- Purchase completion is a better relevance signal than clicks — aligns model to business objective
- Offline metrics (MAP/MRR) predicted online conversion improvement
- LTR outperforms rule-based ranking even with a simple logistic regression model