What is learning-to-rank in recommender systems?

Learning-to-rank trains models to optimize the ordering of items instead of predicting absolute values.

LTR techniques directly optimize ranking quality using approaches like pointwise (predict score), pairwise (compare item pairs), and listwise (optimize full ranking list). Loss functions such as LambdaRank or ListNet are designed to improve ranking metrics like NDCG. It is widely used in search engines and recommendation feeds.