Pairwise Ranking

With implicit feedback there are no ratings to predict — only the fact that someone clicked this and not that. Pairwise ranking takes that fact at face value: instead of learning a score for each item in isolation, it learns from comparisons, training the model so that an item a user engaged with scores above one they did not.

What the model actually optimises
Each training example is a triple — a user, a positive item, and a sampled negative — and the loss depends only on the gap between the two scores. Bayesian Personalized Ranking is the canonical form, maximising the probability that the positive outranks the negative; RankNet applies the same idea with a neural scorer. Absolute score values are free to drift, because only the ordering is penalised.

Why it suits recommendation

• No negative labels required — unobserved items are sampled as negatives, the only workable reading of an implicit matrix in which almost everything is unobserved.
• It optimises what the metric measures — a pointwise regression can halve its squared error without moving a single item up the list.
• Negative sampling is the real lever — uniform negatives are weak; popularity-weighted or in-batch negatives sharpen the model, and sharpen its popularity bias too.

Its limits
A pair loss treats every comparison as equally important, so swapping the items at ranks 400 and 500 costs what swapping ranks 1 and 2 costs — the gap listwise objectives like LambdaRank exist to close. BPR against implicit ALS on MovieLens remains the standard back-to-back look at what a pair loss actually buys.