ranknet对比listnet

The ListNet method grows on the bases of RankNet, they both employ the Cross Entropy function as a loss function and Gradient Descendant as algorithm to train a Neural Network Model. While the ListNet uses document list as instances, RankNet uses document pairs.

We investigated why the listwise method ListNet can outperform the pairwise methods of RankNet, Ranking SVM, and RankBoost.

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-2007-40.pdf

1.for the pairwise approach the number of document pairs varies largely from query to query

2.The pairwise approach actually employs a ‘pairwise’ loss function, which might be too loose as an approximation of the performance measures