• The ranking stage of the Retrieval Ranking Funnel: the candidate set is small (~hundreds), so we can afford a heavy model with cross features. Question: how do we train a model to order items?
• Key insight: ranking quality is about relative order, not absolute scores. Getting the top items in the right order matters; exact score values do not.
• Three families differ by how directly the loss targets order:

1. Pointwise Ranking — predict an absolute score per item, then sort. Ignores order.
2. Pairwise Ranking — penalize wrongly-ordered pairs. Targets relative order, but treats all pairs equally.
3. Listwise Ranking — optimize the whole-list metric (NDCG). Weights the top most.

The progression:

• Pointwise ignores order → pairwise optimizes relative order but weights all pairs equally → listwise optimizes the true list metric and weights the top most. Each fixes the previous one's blind spot.

Metric

• NDCG is the default ranking metric (position-discounted, graded relevance, normalized). Listwise methods optimize it directly.

References