5

I've been reading papers on pairwise ranking and this is what I don't get:

what is the difference in the training/testing data between pointwise and pairwise ranking? This is the paper that I have been reading: http://www.cs.cornell.edu/people/tj/publications/joachims_02c.pdf

In there, it says that a data point in pairwaise ranking is an inequality between two links:

[line] .=. [inequality between two links, which is the target] qid:[qid] [[feature of both link 1 and 2]:[value of 1 and 2]] # [info]

RankLib, however, does support pairwise rankers like RankNet and RankBoost, but the datapoint format that it uses it's that of pointwise

[line] .=. [absolute ranking, which is the target] qid:[qid] [feature1]:[value1] [feature2]:[value2] ... # [info]

Is there something I am missing?

Martin Klosi
  • 3,098
  • 4
  • 32
  • 39

1 Answers1

12

Point wise ranking is analogous to regression. Each point has an associated rank score, and you want to predict that rank score. So your labeled data set will have a feature vector and associated rank score given a query

IE: {d1, r1} {d2, r2} {d3, r3} {d4, r4}

where r1 > r2 > r3 >r4

Pairwise ranking is analogous to classification. Each data point is associated with another data point, and the goal is to learn a classifier which will predict which of the two is "more" relevant to a given query.

IE: {d1 > d2} {d2 > d3} {d3 > d4}

Raff.Edward
  • 6,404
  • 24
  • 34