Approaches

1. Pointwise

• Treat ranking as a regression or classification problem.
• Each item gets a relevance score (e.g., “click = 1, no click = 0”).
• Predict scores, then sort by them.
• Pros: Simple.
• Cons: Ignores relative comparisons.

2. Pairwise

• Train on pairs of items: for a given query, item A should be ranked above item B.
• Loss function penalizes incorrect orderings.
• Example: RankNet, LambdaRank, LambdaMART.
• Pros: Directly models preference comparisons.
• Cons: More complex, scales quadratically with pairs.

3. Listwise

• Look at the entire ranked list at once.
• Optimize a loss that depends on the whole permutation (e.g., NDCG, MAP).
• Example: ListNet, ListMLE, and models using Plackett–Luce.
• Pros: Matches evaluation metrics better.