learnedtorank

Learned to rank (LTR) refers to a family of machine learning methods designed to train models to order items in response to a query according to relevance. Unlike traditional classification or regression, LTR optimizes the ordering of a list of results and is widely used in web search, recommendation, and question-answering systems.

Early work introduced ranking models based on pairwise comparisons such as RankSVM, which learns a scoring function to preserve the order of item pairs. Later approaches explored neural networks (RankNet) and listwise objectives (ListNet, ListMLE). A major development was LambdaRank and its gradient-boosted successor LambdaMART, which use differentiable surrogates for ranking metrics and scale to large data sets.

Common formulations include pointwise, pairwise, and listwise objectives. Pointwise treats each item independently with a relevance score; pairwise uses preference pairs; listwise optimizes entire permutations with respect to evaluation metrics such as NDCG or ERR. Real-world systems often combine LTR with gradient boosting, neural networks, or factorization models and may use libraries such as RankLib, XGBoost with ranking objective, LightGBM, or CatBoost.

Evaluation relies on ranking metrics like NDCG, MAP, or ERR, typically measured on held-out data or via online A/B tests. LTR models require labeled data with graded relevance judgments, which can be constructed from click logs, editorial judgments, or synthetic data; care is needed to control bias and ensure generalization.

Limitations include data requirements, potential bias in judgment data, and computational complexity. LTR remains a core technique for optimizing user experience in search engines and recommendation systems, and it continues to evolve with advances in deep learning and differentiable ranking losses.