L2r

Learning to rank (L2R) refers to a family of supervised machine learning techniques aimed at constructing ranking models that order items in response to a query or context. In information retrieval, the objective is to learn a scoring function s(q,d) that assigns higher scores to more relevant documents d for a given query q, so that a ranked list is produced by sorting by score. Training data consists of query-item pairs with relevance judgments or implicit feedback. Features describe the relationship between query and item, such as textual similarity, term statistics, popularity signals, or user signals. The model is optimized to maximize ranking quality as measured by metrics such as NDCG, MAP, or precision at k.

Approaches are broadly categorized as pointwise, pairwise, and listwise. Pointwise methods treat the problem as regression or classification predicting a relevance score for each item. Pairwise methods learn a function that prefers relevant items over less relevant ones, using losses like hinge or logistic on pairs (RankSVM, RankNet). Listwise methods optimize across entire ranked lists; notable examples include LambdaMART (gradient-boosted trees with a LambdaRank objective) and ListNet.

Applications include web search, e-commerce product search, and recommender systems. Neural and transformer-based models are increasingly used for re-ranking: taking an initial candidate list and reordering with deep architectures (e.g., BERT-based rankers). Evaluation typically uses held-out data with graded relevance and metrics such as NDCG, MAP, and precision@k. Training can rely on labeled datasets such as LETOR or on online signals via online learning to rank.

Challenges include data biases, efficiency for large-scale catalogs, cross-domain generalization, and interpretability. L2R remains a central toolkit for modern search and ranking systems.