Learning-to-Rank
Best for: Ranking search/recommendation results
How it works
$$L=\sum_{i\succ j}\log\!\bigl(1+\exp(s_j-s_i)\bigr)$$Treats ranking as supervised learning over query-document features using pointwise, pairwise, or listwise losses. Pairwise losses (RankNet) apply logistic loss to score differences for a preferred pair $i\succ j$, $L=\log(1+\exp(s_j-s_i))$, while listwise methods such as LambdaMART or ListMLE weight gradients by ranking-quality gains like NDCG. The score $s=f_\theta(x)$ is typically produced by a gradient-boosted tree or neural ranker trained on millions of judged query-document pairs.
Common fields
Search engines · ads · marketplaces