LambdaRank
A pairwise ranker trained with a smooth loss will happily spend as much effort fixing an inversion between rank forty and forty-one as one between first and second. LambdaRank corrects that imbalance with an unusual move: leave the loss alone and rewrite the gradient.
Rewriting the gradient
In a neural pairwise ranker such as RankNet, the gradient on one item is the sum of forces from every item it was compared against. LambdaRank multiplies each force by the absolute change in NDCG that would follow from swapping that pair. A pair straddling the first two positions yields a large lambda; the identical pair deep in the list yields almost nothing. No differentiable surrogate for the metric is constructed — the weighting is applied straight to the update, and empirically those updates drive NDCG to a local optimum.
Why it matters here
Recommendation labels are implicit — a click, a play, a purchase — and overwhelmingly absent, while only the first few slots of a list are ever examined. Metric-weighted gradients put the model's capacity exactly there, which is why pointwise scoring of each item on its own loses on top-k metrics even when its probability estimates are better calibrated.
The name to keep straight
LambdaRank is a gradient definition, not a model. Drop those lambdas into gradient-boosted regression trees and you have LambdaMART; hand them to a neural scorer in the re-ranking stage and you have the objective most production ranker training uses, available directly in LightGBM and TensorFlow Ranking.
LambdaRank is a pairwise learning-to-rank method that scales each pair's gradient — its lambda — by the change in NDCG that swapping the two items would produce, so the model is pulled hardest toward fixing mistakes near the top of the list. It makes a position-sensitive, non-differentiable metric trainable by gradient descent, and supplies the gradient definition that LambdaMART plugs into gradient-boosted trees.
Picture marking exam papers when only the top three results will ever be read out. Fixing a mistake between the first and second paper matters. Fixing one between the fortieth and forty-first changes nothing anyone notices.
LambdaRank teaches software to spend its effort the same way. It learns by comparing two things at a time and asking which should come first — but it weights each comparison by how much getting it right would improve the part of the list you actually look at. Mistakes near the top are corrected hard. Mistakes far down are nearly ignored. What comes out is an ordering tuned to the few slots on your screen rather than to the whole catalogue.