Pairwise Ranking
- A Learning to Rank family that looks at pairs of items.
- For a pair (A, B) where A is more relevant than B, the model should score A > B.
- Loss penalizes inversions (pairs in the wrong order).
- Classic method: RankNet — a logistic loss on the score difference
s_A - s_B.
Why better than pointwise
- The loss directly targets relative order, which is what ranking is about. It does not waste effort on absolute calibration like Pointwise Ranking.
Weakness
- Treats all inversions as equally bad. But an inversion at ranks 1-2 hurts the user far more than one at ranks 499-500. Pairwise loss does not know that. Listwise Ranking fixes this.