Learning to rank is the branch of machine learning concerned with ordering items, rather than classifying or scoring them in isolation. The classic setting is web search: given a query and a set of candidate documents, produce an ordering that puts the most relevant results at the top. What matters is the order, not the absolute predicted score, because users mostly look at the first few results.
The field is usually divided into three approaches. Pointwise methods predict a relevance score for each item independently and sort by it, treating ranking as ordinary regression or classification. Pairwise methods, such as the RankNet model introduced by Chris Burges and colleagues at Microsoft in 2005, learn from comparisons: document A should rank above document B, optimizing a function whose score differences match those preferences. Listwise methods go further and optimize a loss defined over the whole ordered list, trying to directly improve ranking metrics like NDCG. A practical complication is that those ranking metrics depend on sorted order and are not smooth, which is why influential methods like LambdaRank and LambdaMART work with carefully designed gradients rather than the metric itself.
Learning to rank powers web search, product search, recommendation feeds, and the candidate-ranking stage of advertising systems. It is the discipline that replaced hand-tuned ranking formulas with functions learned from large collections of human relevance judgments and behavioral data.
For a business reader, learning to rank is the quiet engine of the consumer internet: nearly every ordered list you scroll, from search results to shopping pages to recommendations, is the output of a model trained to put the right things first.