Learning to Rank journey at GetYourGuide: Our Logbook #bbuzz
summary
GetYourGuide builds an LTR pipeline on Databricks + Spark, collects 30,000 judgments from internal experts (0–3 scale), engineers query‑document, business (clicks, bookings, CTR/CR) and item features, trains models with RankLib / LambdaMART, and evaluates by NDCG@10. The system iterates feature/model versions but the best offline model never ships: key queries like “Eiffel Tower skip the line” rank poorly, judgments are nearly binary and noisy, and business signals dominate, so data‑driven judgments and cleaner query intent are required before production.
AI-written summary · about
chapters (10)
- 0:00Introduction and GetYourGuide overview
- 1:51Why learning to rank at GetYourGuide
- 4:21Project scope and first iteration
- 7:10Collecting human judgments for training
- 10:43Feature engineering and data pipeline
- 13:38Building the learning to rank pipeline
- 17:15Training and validating the model
- 20:51Offline evaluation and lessons learned
- 28:42Redefining relevance and future steps
- 31:35Conclusion and Q&A
related talks
47mLearning to rank search results
30mJourney Towards Deep Learning for Ranking
29mFrom Trees to Transformers: Our Journey Towards Deep Learning for Ranking
38mLearning to Rank: where search meets machine learning #bbuzz
25mBetter search relevance using Learning to Rank at mobile.de
31mSponsored Session: From Clicks to Rankings: Building a Learning-to-Rank Pipeline with…
26mTeaching Ruby to Rank: A Search Pipeline Story
24mLearning "Learning to Rank"