Learning to Rank

A model can predict each item's click probability perfectly well and still produce a bad list, because what the reader sees is an order, not a set of numbers. Learning to Rank is the family of supervised methods that train on the order itself.

Three ways to frame the loss

• Pointwise — score each item on its own against a label, then sort. Simple, reuses any regressor, and blind to the fact that only the relative order of neighbouring items matters.
• Pairwise — learn which of two items should come first, as RankNet and LambdaMART do. This matches the comparison a user implicitly makes on screen.
• Listwise — treat the whole list as one training example and optimise a list-level objective, such as ListNet's distribution over permutations or a smooth approximation of NDCG.

What the recommendation setting adds
In search, relevance labels come from human judges. Here they come from the log: a click is a weak positive, an impression without a click is an ambiguous negative, and an item never shown carries no label at all. That moves the hard part from the loss function to the data — position bias has to be corrected, negatives have to be sampled deliberately, and offline NDCG gains repeatedly fail to reproduce as clicks online.

Where it runs
In the re-ranking stage, over the few hundred candidates retrieval already narrowed down; the full catalogue is never ranked. LightGBM, XGBoost and TensorFlow Ranking all ship these objectives, so the engineering question is rarely which library and almost always which labels you can trust.