Learning-to-Rank
Learning-to-Rank (LTR) is a family of supervised machine learning techniques that train models to produce optimal orderings of items — documents, products, or recommendations — relative to a query or context. Rather than predicting a single label or value, the model learns a scoring function whose output, when sorted, maximises a ranking quality metric such as NDCG or MAP. Approaches are grouped into three paradigms: pointwise (treat each item independently), pairwise (compare item pairs), and listwise (optimise the full ranked list directly).
Search quality is a primary competitive lever for any platform handling product discovery, document retrieval, or content feeds, and LTR is the standard production technique behind it. As companies move from keyword-based BM25 baselines to two-stage retrieval + re-ranking architectures, engineers who can train, evaluate, and deploy LambdaMART or neural ranking models are in high demand at firms like Google, Meta, Amazon, Spotify, and Booking. LTR skills are also increasingly required in RAG pipelines, where re-ranking retrieved chunks before sending them to an LLM directly improves answer quality.
🎓 Courses
Learning to Rank in Web Search — Comprehensive Blog Course
by hav4ik
Published September 2024, this ~99-minute deep-dive covers pointwise/pairwise/listwise methods, LambdaRank, Unbiased LTR, and hands-on MSLR-WEB30K experiments with LightGBM. One of the most current and comprehensive free resources available.
Learning to Rank for Information Retrieval — University of Edinburgh Lecture (TTDS 2023)
by University of Edinburgh (Text Technologies for Data Science)
Rigorous academic lecture notes from 2023 covering the theoretical foundations of pointwise, pairwise, and listwise LTR, RankSVM, RankBoost, and evaluation with NDCG. Excellent companion to hands-on tutorials.
Elasticsearch Learning to Rank — Official Introduction
by Elastic
Covers end-to-end LTR in production: training an XGBoost LambdaMART model, deploying via Eland, and using it as a two-stage rescorer in Elasticsearch 8.15+. Practical and directly applicable to real search systems.
Learning to Rank — OpenSearch Plugin Documentation
by OpenSearch / AWS
Official docs for the OpenSearch LTR plugin, covering feature stores, XGBoost/RankLib model training, and deployment in AWS OpenSearch Service. Essential reading for practitioners building search on AWS.
Learning to Rank with XGBoost — Official Tutorial
by XGBoost Contributors
Official hands-on tutorial for XGBoost's rank:pairwise, rank:ndcg, and rank:map objectives with code examples. Covers query groups, evaluation metrics, and the LambdaMART implementation used in most production search systems.
📖 Books
Learning to Rank for Information Retrieval
Tie-Yan Liu · 2011
The canonical academic reference on LTR — the only comprehensive textbook covering all three LTR paradigms, theoretical guarantees, and evaluation frameworks. Still the standard citation in research papers and the best starting point for understanding the field's foundations, despite its publication date.
🛠️ Tutorials & Guides
Learning to Rank: A Complete Guide to Ranking using Machine Learning
Published August 2024, this illustrated guide explains all three LTR paradigms, LambdaRank, LambdaLoss, and evaluation with MAP and NDCG. Accessible yet rigorous — a good starting point before diving into papers or production tutorials.
Learning to Rank with Elasticsearch — Jupyter Notebook
End-to-end runnable notebook: generate training data, train an XGBoost ranker, upload via Eland, and measure NDCG improvement before and after. The fastest hands-on path from zero to a deployed LTR model.
Learning to Rank for Amazon OpenSearch Service
Step-by-step guide for building an LTR pipeline on AWS using click-through behavioural data, XGBoost/RankLib model training, and the OpenSearch LTR plugin. Practical reference for teams already on AWS infrastructure.
Learning resources last updated: June 18, 2026