Sponsored Session: From Clicks to Rankings: Building a Learning-to-Rank Pipeline with…
summary
An OpenSearch pipeline that gathers client-side click and query events using User Behavior Insights (UBI/UBIJS), joins them by query_id into query and event indices, and creates judgment lists using click-engineered (CE) and LLM-based (LRM) methods. It runs quantitative experiments in Search Relevance Workbench (SRW) with pointwise, pairwise and hybrid setups, trains an XGBoost ranking model externally, registers the model, and performs in-cluster inference with Learning to Rank (LTR) to rescore production results without added latency.
AI-written summary · about
chapters (6)
- 0:00Learning-to-rank pipeline
- 7:26Measuring search performance
- 14:50Search Relevance Workbench
- 22:20Learning-to-rank implementation
- 28:06Optimization cycle
- 30:12Q&A
related talks
26mTeaching Ruby to Rank: A Search Pipeline Story
47mLearning to rank search results
38mLearning to Rank: where search meets machine learning #bbuzz
39mIntroducing OpenSearch’s Search Relevance Workbench!
41mIntegrating Realtime User Behavioral Embeddings Into OpenSearch
40mLevel Up: Everything you need to know about search, ...
40mAgentic Relevance Tuning With OpenSearch: Faster, Easier, and Higher Quality Search at Scale
43mLearning to Rank for Reddit Search: A Project Retro