LambdaMART

Stub. Created as a placeholder — expand with vault-sourced content.

Definition

LambdaMART (Burges, 2010) is the dominant production Learning to Rank algorithm. It combines MART (Multiple Additive Regression Trees = gradient-boosted decision trees) with LambdaRank gradients, where the gradient of each pairwise swap is weighted by its impact on a ranking metric (typically NDCG).

Vault references (existing coverage)

• Learning to Rank — has a full LambdaMART section + LGBMRanker code
• How LambdaMART Works — Doug Turnbull explainer (lambda gradients + MART)
• Part 1 - Learning to Rank for E-Commerce Search at Otto — production use
• Building a Better Search Engine for Semantic Scholar — LightGBM + LambdaRank reranker
• Learn-to-Rank with OpenSearch and Metarank — Metarank implements LambdaMART as an external re-ranker
• Hybrid Search and Learning-to-Rank with Metarank — LambdaMART for multi-retriever fusion
• Metarank - Personalized Ranking That Actually Reads Your Clicks — LambdaMART via YAML config

TODO

• Decide whether this stays a standalone note or stays folded into Learning to Rank (currently an alias there).
• Link to implementations: LightGBM, XGBoost, CatBoost, RankLib.

Implementations & Serving

• Libraries: LightGBM, XGBoost, CatBoost (all handle missing feature values natively), RankLib
• Serving: Metarank — open-source service that trains and serves LambdaMART as a secondary re-ranker

Related

• Learning to Rank · Ranking Objectives · NDCG · RankNet · Metarank