Home Glossary Item Learning-to-Rank
« Back to Glossary Index

Learning to rank, sometimes called machine-learned ranking, is a significant aspect of machine learning that focuses on constructing models used to sort items into a specific order. This technique is largely used in applications where the items’ order matters such as e-commerce recommendation systems, document retrieval in search engines, and information retrieval (IR) where documents need to be ranked by relevance to a particular query.


Learning to rank involves training an algorithm on a set of items with predetermined rankings. The objective is to deduce the inherent patterns or relationships that influence the ordering so that the model can then automatically rank a new set of items accordingly. Three primary approaches to this problem exist- Pointwise, Pairwise, and Listwise approaches. While Pointwise models treat each item individually and model the problem as a regression or classification task, Pairwise methods focus on relative order between pairs of items. In contrast, Listwise methods model the entire list of items and their ranking directly.

There are numerous algorithms for achieving effective learning to rank systems. A common choice is the use of gradient boosting machines, with LambdaMART being a widely-used variant. Other popular solutions may include RankNet, LambdaRank, and RankBoost. Irrespective of the algorithm used, the ultimate objective remains the same – to create a ranking model that is able to provide an ordering that is as close as possible to the ideal or true ranking of the items.

« Back to Glossary Index