A pairwise transform converts learning-to-rank judgments into binary-classification instances. For each query, it compares documents with different relevance grades, subtracts their feature vectors, and labels the difference +1 or -1 according to which document is more relevant. section 10.4.2
The transformation proceeds as follows:
- Group judgments by query and compare every pair with different grades. section 10.4.2
Search-ranking example
For “social network,” subtracting Network’s [3.101, 1.443] from The Social Network’s [4.483, 2.100] produces a positive training example; reversing the pair produces the corresponding negative example.
