pairwise ranking
**Pairwise ranking** learns **from item comparisons** — training models to predict which of two items should rank higher, directly learning relative preferences rather than absolute scores.
**What Is Pairwise Ranking?**
- **Definition**: Learn which item should rank higher in pairs.
- **Training Data**: Pairs of items with preference labels (A > B).
- **Goal**: Learn function that correctly orders item pairs.
**How It Works**
**1. Generate Pairs**: Create pairs from ranked lists (higher-ranked > lower-ranked).
**2. Train**: Learn to predict which item in pair should rank higher.
**3. Rank**: Use pairwise comparisons to order all items.
**Advantages**
- **Relative Comparison**: Directly learns ranking order.
- **Robust**: Less sensitive to absolute score calibration.
- **Effective**: Often outperforms pointwise approaches.
**Disadvantages**
- **Quadratic Pairs**: O(n²) pairs for n items.
- **Inconsistency**: Pairwise predictions may be inconsistent (A>B, B>C, C>A).
- **Computational Cost**: More expensive than pointwise.
**Algorithms**: RankNet, RankSVM, LambdaRank, pairwise neural networks.
**Loss Functions**: Pairwise hinge loss, pairwise logistic loss, margin ranking loss.
**Applications**: Search ranking, recommendation ranking, information retrieval.
**Evaluation**: Pairwise accuracy, NDCG, MAP, MRR.
Pairwise ranking is **more effective than pointwise** — by learning relative preferences directly, pairwise methods better capture ranking objectives, though at higher computational cost.