Learning to Rank

A family of supervised ML techniques that learn a ranking function from labeled query-result pairs, replacing hand-tuned scoring formulas with a model trained on relevance data.

Learning to Rank (LTR) frames ranking as a supervised problem: given a query and a candidate list, output an ordering. Three flavors: pointwise (score each item independently), pairwise (predict which of two items is more relevant — RankNet, LambdaRank), and listwise (optimize the full ordering directly — LambdaMART, ListNet).

In ecommerce, LTR shines as the re-rank stage on top of BM25 and vector retrieval. Features feed in: text similarity scores, click-through rates, conversion rates, recency, in-stock status, price, margin, brand authority. The model figures out the weighting better than a hand-tuned linear formula.

You need labels: either explicit (human raters scoring (query, product) pairs) or implicit (clicks and purchases as positive signals, with position-bias correction). Most stores start with implicit labels at scale and only invest in explicit labels for evaluation sets.

Related terms

Re-ranking

A second-stage scoring pass that takes the top-N results from a fast retriever (BM25 or vector) and re-orders them with a more expensive, higher-quality model.

Cross-Encoder

A transformer that scores query and document together in a single pass — slower than bi-encoders but much more accurate, used as a re-ranker on the top-N candidates.

NDCG

Normalized Discounted Cumulative Gain — a ranking quality metric that rewards relevant results highly placed in the result list and discounts those buried below the fold.

Click-Through Rate (CTR)

The percentage of search-result impressions that result in a click — the primary engagement signal for measuring search ranking quality.

Browse all glossary terms