From Trees to Transformers: Our Journey Towards Deep Learning for Ranking
summary
GetYourGuide replaces an XGBoost ranker with a production PyTorch transformer over nine months, keeping throughput and latency acceptable after tuning CPU threads and docker size (initial image >4 GB) and solving multi‑GPU embedding bugs. The team runs ~50 offline models and 10+ live A/B tests, stages work into technical‑parity, business‑parity and business‑win milestones, adds semantic similarity and trainable embeddings, and finally ships a deep & cross + transformer stack that produces measurable booking and revenue uplift.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:11Why ranking matters
- 4:00Learning to rank
- 6:08Tree-based to deep learning
- 9:17Deep learning challenges
- 11:25Milestone one parity
- 12:31Milestone two business priority
- 13:34Milestone three transformer
- 14:58Milestone four semantic similarity
- 16:01Model architecture
- 16:46Ranking transformer deep dive
- 20:40Key takeaways
- 21:45Project milestones
- 22:30Closing remarks
- 22:56Q&A
related talks
30mJourney Towards Deep Learning for Ranking
29mLearning to rank with the Transformer
38mLearning to Rank journey at GetYourGuide: Our Logbook #bbuzz
34mTransformers End to End Experience training...
23mUnveiling the Journey of NLP
25mTransformers.js: Building Next-Generation WebAI Applications
20mPyTorch x Transformers Journey
29mRanking hotel images using deep learning